Beruflich Dokumente
Kultur Dokumente
VOLUME LXXX
AVARDED THE DAVID A. WELLS PRIZE FOR THE YEAR 1941-42 AND
PUBLISHED FROM THE INCOME OF THE DAVID A. WELLS FUND. THIS
PRIZE IS OFFERED ANNUALLY, IN A COMPETITION OPEN TO SENIORS
OF HARVARD COLLEGE AND GRADUATES OF ANY DEPARTMENT OF
HARVARD UNIVERSITY OF NOT MORE THAN THREE YEARS. STANDING,
FOR THE BEST ESSAY IN CERTAIN SPECIFIED FIELDS OF ECONOMICS
BY
PAULANTON _SAMUSN
PROFESSOR OF ECONOMIC-S.: "" _ ,. ,
'iASSAc'HU's' rra ' 05TECHNOLOGY
@@
@;
Mathematics is :: Language
J. WILLARDGrans
CAMBRIDGE
HARVARD UNIVERSITY PRESS
Cowman, {947
BY THF PRESIDENT AND FELLOWS OF HARVARD COU-EGP
sun 674-33093
PRINTED IN rm-: UNITED auras or AMERICA
To
MY PARENTS
PREFACE
1'r-IE ORIGINAL version of tl1is book sub111itted to the David A.
vVells Prize Committee of Harvard University in 1941 carried the
subtitle, "The Operational Significance of Economie Theory.'' At
that time most of the material presented was already several years
old, having been conceived and \vritten primarily in 1937. Further
delay in publication 11as been necessary because of the \var, and
because of the addition of supplementary treatise-Iike n1aterial
going beyond the original conception of the work as indicated by
its subtitle.
Because of the pressure of war work I have not been able to
do full justice to the literature of the last few years, nor even to
include all of the developments of n1y own thinking. Fortunately,
the passage of time has dealt kindly with the analysis contained
here, and where it abuts upon the topics treated in Professor
Hicks's masterly Valite and Capital, the similarity in point of view
11as been reassuring.
My greatest debt is to Marion Crawford Samuelson whose con-
tributions have been ail too many. The result has been a vast
mathematical, economic, and stylistic improvement. Without her
collaboration the book \Vould literally not have been \vritten, and
no perfunctory uxorial acknowledgment can do justice to her aid.
Nor can tl1e quaint modem custom of excluding the value of a \vife's
services fron1 tl1e national income condone her exclusion from the
title page.
My thanks for prolonged stimulation 0\1er rr1any years must go
out to Professors Schumpeter, Leontief, and E. B. Wilson, while
each of a legion of Harvard graduate students has left his mark
upon what follows. The reader will note my dependence upon the
sterling contribution to Welfare Economies of Professor Abran1
Bergson. Grateful acknowledg1nent is made to the Social Science
Research Council and to the Society of Fellows of Harvard Uni-
versity for the opportunities they provided for pursuit of inde-
pendent research, and to the Department of Economies of Harvard
University for their courteous acceptance of the wartime delays
in publication.
vu
vm PRE FACE
2xy
r=-:
V2x2 \/y2
Discontinuous Case 80
External Conditions of Equilibrium. 81
Summary . . . . . . 87
V. THE PURE THEORY OF CONSUMER'S BEHAVIOR 90
Evolution of Utility Concept. . 90
Progression1nMathematical Thought. 92
The Demand Functions as a Goal . 96
Conditions of Equilibrium 97
Displacement of Equilibrium. . 100
Meaningful Theorems. . 107
Conclusion . . . 116
A Note on the Demand for Money. 117
Qualications Introduced by Uncertainty . 122
VI. TRANSFORMATIONS, COMPOSITE COMMODITIES, AND RA
TIONING . . . 125
LogarithmicTransformationsand Elasticities.125
General Transformation OfIndependent Variables. 129
Transformation of Dependent Variable. . 133
Transformation
ofPrices... . 135
Demand for a Group of Commodities . . .141
The General Problem of Composite or Aggregate
Commodities .144
The Economic Theory of Index Numbers . . 146
Present Formulations of Index Numbers . 156
Pure Theory of Choice under Rationing . 163
VII. Sons SPECIAL ASPECTS OF 1111-:THEORY or Con
SUMER's BEHAVIOR. . 172
The Cardinal Measure of Utility . . 173
The Assumption of Independent Utilities . . 174
Complementarity. . . 183
Constancy of the Marginal Utility of Income. . 189
Why Consumers Surplus 1sSuperuous . 195
The Many Forms of Consumer's Surplus . . 197
VIII. WELFARE ECONOMICS . . 203
The Social Welfare Function. . 219
Mathematical Analysis . 229
Production Conditions . 230
CONTENTS xi
Pure Exchange Conditions . 236
Interpersonal Optimal Conditions . 243
New Versus Old Welfare Economics . 249
Conclusion . . 252
PART II
IX. THE STABILITY OF EQUILIBRIUM: COMPARATIVESTATICS
AND DYNAMICS . . 257
Introduction . . 257
Comparative Statics . . 258
Stability and Dynamics . . . . 260
The Stability of Multiple Markets . . 269
Analysis of the Keynesian System . . 276
X. THE STABILITY OF EQUILIBRIUM: LINEAR AND NON
LINEAR SYSTEMS . 284
Introduction . 284
Functional Equations and Stationary Solutions . 286
Linear and Nonlinear Systems 288
The Nonlinear Di'erentialEquation InOne Vari
able . . 288
Example: Logistic Law . . 291
The Problem of Higher-Order Stability. . 294
An Example of One-sided Stability-Instability:
Malthusian and Optimum Population Theories. 296
Systems of Equations m n" Variables. 299
DISPLACEMENTOF EQUILIanu
It 15easy to show mathematncally how the rates of change of
our unknownc thh respect to any parameter, say on.may be com
puted from our equtllbnum equations As a matter of notation let
__g'(al. . "-)
6 Bar g1(m . .a )
EQUILIBRIUM AND COMPARATIVE STATICS 13
stand for the rate of change of the ith variable with respect to the
parameter al, all other parameters being held constant. It is
necessary because of the ambiguity of the conventional notation
of partial differentiation to make sure which variables are being
held constant.
Such partial derivatives must be evaluated as of a given value
of the set of parameters, and hence as of a corresponding set of
values of our dependent variables. Consider the initial position
(ale) ' ' ' ) amo),
1 _. 1 _, ... = .. ,
f2) (6&1)
6x1 +fr, ( ad]
6x2 )0 + +f2,1 (336,.
al )0 f l
2 __ 2 __ ... : __ .. a! ,
f1 (a)
6x1 "He (a
6x2 ) + H" 2 (xn
aa, ) _ f 2
. _ . (6)
n(Gal)
f-l'l 2.x."+f22"(ray
aal + +fn "(axnw
al fl
where
i afi(x10, , , , , xno, (11,. . ., @
f : ax,- ,
with all other variables and parameters held constant. Similarly,
,. fi(x1. ' ' ' , xnoy (1109 ' ' ' , amo) .
fa] = 60]
2In matrix terms this is [j,-"][xj/m] = [" fai]
L} FOUNDATIONS OF ECONOMIC ANAL}SIS
Be rt noted tlnt the numernml \alues of these part: 11dematwes
are fully determmed at the (.qmllbnum pomt la question Thus
we have 72linear equatlons wzth constant coefcrents m n unknou ns
[Gin/am)" ext/amp] The solution xalues of these wrllde
pend upon the values of the coeferents thus the partral denva
twes relating our dependent V'mables md parameters are deter
mmed by the structural properttes of our equzhbrrum system
Since (6) represents linear equations their solution for non
srngular cases may be represented m the familiar detenmnantal
form
0 if. A&
(%:) = 'A (le=1 n) (7)
here
j! fll f!
f: fz. fx.
A = If | =
f2 1:. fx.
'md A g 1ndtcates the COflCtOI
of the element of the zth row and the
kth column Or nntruc terms
[ai]
6C1] = UT'Ef-J (8)
ILLusraAnvn TAXPRoan
For concreteness let us apply our mal) 515to two snnple cases
Consxdera rm thh '1gmen demand curve relatmg puce and out
put and '1 gtxen productxon cost schedule relatrng total cost and
output Suppose m addrtlon that the output ol the rm rssubject
to '1tax of t dollars per umt The prot of tl e rm may be wntten
r = tdtzreveme wr'u'lprufrum -ccfrcapvynaw.
= xp(x) C(zx) -h
where
xp(x) = total revenue m function of output
C(x) = lowest total productnon cost at wlnch each output can
be produced
tx = total tax payment
EQUILIBRIUAI AND COMPARATIVE STATICS 15
It is clear that for any given tax rate, say t, the rm will decide
upon some given output to be produced and sold; i.e.,
96"= g(t). (9)
where the functional relationship g corresponds to the functions
in (2). However, we cannot leave the matter in so indenite a
state. We wish to know more than that there exists an equilibrium
output for each tax rate. What is the nature of the dependence
of our variable upon the tax rate regarded as a parameter? Will
an increased unit tax result in a larger or smaller output? It is a
poor theory indeed which will not answer so simple a question.
Let us see if we can arrive at an answer to this through the formula
tion of the conditions of equilibrium.
It is in general assumed that a rm will select that output which
will maximize its net revenue. That is to say, our equilibrium
value for output will emerge as a solution of a simple maximum
problem. Specically, it is necessary 3 for a regular maximum of
prot with respect to x, as of a given tax rate, that
7r(x, t)
x : 0 10
62w(x, t) < 0 ( )
ax2 '
;; (mx) - can t = 0, (
where it is assumed for simplicity that the inequality in (10).is
realized for all values of the variable concerned. Now equation
(11) corresponds to our equilibrium set (1), which in thlS case con
tains but one equation due to the fact that we have only the value
of one unknown to determine.
See Mathematical Appendix A, sec. 1.
16 FOUNDATIONS OF ECONOMIC ANALYSIS
@@)"
6x a _ (?2)
a _ 3:12.
a
(18)
seia efdo: _o
'
By simple substitution we get
6_Q
(9E7) - ""
6a _ _ _ S"
(19)
5,212
(
a _- _ _
D ,.
- S
Now we know that D/a >. 0 by denition of our parameter of
shift. Thus, (ax/600: 0 depending upon whether 5 D/x.
At rst sight this seems merely to put off the dilemma, to replace
one equation by another. But if we examine the kind of market
This is treated in greater detail in chap. ix below.
l FOUNDATION-S OI ECONOMIC ANAILSIS
under conSIderatton, It wdl appear that the mere fact that the
market 18m stable equllnbrlum m the lmtlal Sltuauon W11!remove
all amb1gu1ty For 1fthe market ISthe famthar Marshalhan con
sumer 5good market stabnhty of equrhbnum by demtlon requnres
that the supply curve cut the demand curve from below (even m
the case of decreasmg cost due to external economles)
Thus,
, D
S > 2935 (20)
Therefore,
61:
(a?) > (21)
Hou ever, the algebralc s1gnof the price change w1llhmge upon
whether there 1sa posntxvelyor negatively mclmed supply curve
For
BP 0 .. , E !
(a) -S (au) (22>
Thus, smce (at/600 ISposmve from (21) (GP/600and 5 must be
of the same Slgn, or
+ _ _ ... _
L_'"'
T-Iowmany of these can now be eliminated as inadmissible on the
)asis of our qualitative knowledge as embodied in the specication
26 FOUNDATIONS OF ECONOMIC ANALYSIS
of the matnx (1)? Ideally, we should hope to be able to rule out
all but one of the possxblecombmat1ons so as to get a unique answer
However, xt would be at least desnrable to be able to determine the
Slgn of at least one of our unknowns, or, what comes to the same
thmg to be able to rule out a partlcular half of all the possnble
combmatlons x L/
So much for our asp1rat10ns, turmng to the exact procedure by
whlch we rule out a comblnat1on, we nd ourselves m for dns
appomtment If we substitute mto equat1on (6) of chapter 11the
sngns of the 1nd1cated elements, we shall be able to rule out '1
cornbmatton If and only If lt leads to a contradnctron 1e 1fxtdoes
not add up to zero, or to a negatzve number as It should
Concretely, this can be seen to mean that we can rule out any
one of the above combmatxons of sxgn m (2) Wthh exactly duph
cates any of the last (n 1) rows of (I), or whlch IS the exact
antIthesm of any of these same rows Otherwxse, one of the last
(1: 1) equations of (6) chapter n could not add up to zero as
reqmred by the present hypothesrs In addntton we may rule out
any comb1nat1onof (2) \Vthl'l exactly duphcates the rst row but
we are not able to rule out its antlthems /o 64-5 :_
All told then we are at most able to rule out by quahtatlve
conSIderatlons only (Zn I) combmatlons out of a grand total of
2" posmble comblnatlons Even for moderate srzes of n thusleaves
us w1th all but a mmute fraction of possnbnhtnes And th1s Is the
largest number that we can rule out on such a hypothesns In
many cases we are not able to rule out even thxs many Thus, If
any two rows our orngmal matrrx have exactly the same Signs,
they wxllboth rule out the same combmat1ons Wthllcannot there
fore be added together for fear of double countmg
It can be seen then that purely qualltatxve consxderatronscan
not take us very far as soon as the srmple cases are left behind
Of course If we are w1llmgto make more ngld assumptlons elther
of a quahtatwe or quantltatlve kmd we may be able to Improve
matters somewhat Ordmanly, the economist 15not m possessuon
of exact quantztatne knowledge of the partial der1vat1vesof his
cqunhbnum condntlons None the less 1f he 15 a good applied
Leonomnst, hc may have demte notrons conccrmng the rclatnc
1mportancc of dtffercnt e'ects. the better hrs Judgment m these
matters the better an economlst he Willbe These notions whach
THEORY OF MAXIMIZING BEHA VIOR 27
are anything but a priori in their original derivation, may suggest
to him the advisability of neglecting completely certain effects as
being of a second order of magnitude. In other words, zeros are
inserted into the matrix of (1). In fact, the so-called method of
partial equilibrium consists of nothing more than a liberal sprinkling
of zeros into the equations of general equilibrium. In the hands
of a master practitioner, the method will yield useful results; if not
handled with caution and delicacy, it can easily yield nonsensical
conclusions.
By simple extensions of the above argument we can show how
the presence of a zero in any row permits that row to rule out four
instead of only two combinations. Similarly, r zeros in a row per
mits that row to eliminate 2""l combinations. As before, there
may be duplications in the eliminating inuence of dierent rows.
It is also apparent that denite knowledge concerning the sign of
any one of the variables will enable us to rule out one-half of the
original number of combinations, or 2 combinations. Because
of duplicating eects, denite knowledge of the signs of the changes
in two of the unknowns will enable us to eliminate less than twice
as many as one unknown; actually, by two known signs we can
eliminate 3(2"2) combinations in all. Denite knowledge of the
signs of k unknowns will permit the elimination of all but 2"""
combinations.
It would be possible to illustrate the above calculus of quali
tative relations by reference to a number of well-known economic
problems. Space will permit brief mention of a few. In the
market illustration of the previous chapter there was considered
the case of a simple Marshallian partial equilibrium market in
volving two unknowns, price and quantity, whose equilibrium
values are determined by the intersection of supply and demand
schedules. We may apply our analysis to the determination of the
changes in our variables resulting in an assumed shift of a nega
tively inclined demand curve. If the supply curve is known to be
positively inclined, the sign matrix can be written in the form
[t :1
Page 17.
28 FOUNDATIONS OF ECONOMIC ANALPSIS
ln th1s case, we are able to rule out the maumum number of com
bmatxons, (2n 1), or three m all Smee there are only {our
combmatnons we are left W1tha umque answer, as 15apparent from
the algebraic analyms of the prev10uschapter If the supply curve
IS assumed to be negatlvely Inchned the sngns m the second row
both become plus, and we are able to rule out only the minimum
number possnble,or two We are left then thh a nal amblgmty
wluch cannot be settled except by quantztatwe knowledge or by
means of varlous stabIhty hypotheses
A more lllummatmg and more difcult example nsthat of the
smphed Keynesmn system, desenbed greater detall m chapter
1x \thhout gomg Into detalls here, It can be stated that thls gives
use to a system of three varlables, Interest rate xl, Income x, and
Investment, x; whose equlllbnum values are determmed by three
equatlons, hqmdrty preference. f , margmal efcnencyschedule,f2,
and propenSIty to consume f Makmg the usual assumptlons
concermng the sngns of the Varlous rst order effects, eg that
mvestment vanes mversely wuth the rate of Interest and that the
effect of mterest on consumption 18of an order of magmtude whxch
can be neglected, ue end up w1th a sngnmatm of the form
+ - 0
(Sign fr) = - + (4)
o - +
We may now consxder a shalt m any one of the schedules e g
1 change m the hqmdxty preference schedule brought about by
some polncy measure If we apply the above calculus we are snll
left wnth three posmble combmatlons for (51gn change m x.)
namely, (+ + +), ( + +),or ( ) ln thecase ofa shllt
of the mammal eiuency schedule, we narrow the ch01cedonn to
two, (+ + +) or ( ---) For a Shift m the propensnty to
Lonsume schedule ue end up mth three chonces, (+ + +).
(+ + ).0r(- -_)
lt mll be noted that 1nnone of the cases are no able to make a
demte statement about even one of the variables In the sccfmd
case, Wthh xsthe most favorable, we can only say that an Increase
m the margmal emency schedule wxllenther ranse Interest, mcomc,
and Investment, or else lou cr all three of them Thrs ISa highly
unsansfactory state of affairs partlcuhrly smce lt seems on the
THEORY OF JlIAXIJlIIZING BEHAVIOR 29
face of it doubtful that an increase in the marginal efciency of
capital should lower interest, income, and investment. But it is
only by bringing in the stability considerations of later chapters
that a more denite, and more reasonable deduction can be made.
In the discussion in chapter ix it is shown that stability hypotheses
will leave us with only one combination for the rst case, ( + +);
for the second case, (+ + +) ; and for the third case the two possi
bilities, (+ + +) or (+ + -). This last ambiguity is irremov
able unless still stronger quantitative assumptions are made.
Numerous other economic examples could be mentioned, but
these may be left to the reader.7 Before returning to the problem
of maximizing behavior, I should like to call attention to the fact
that the qualitative calculus is not invariant under transformation
of the variables.
where X stands for the arguments (x1, , x), and a stands for
the arguments (ax, , a)
If z represents an absolute maxzmum wrth respect to all ad
m1551blevalues of our Independent variables, then
f(X. c2)< f (X. ) (8)
On the other hand, z0may merely represent a manmum relatwe to
all c's lymg In some restricted nerghborhood of the pomt (X. a)
We know from Mathematlcal Appenduc A, Section II that m
order for z to enjoy a relative maxxmum 1t rs necessary that
6
: 0 = f,,(x1, ; xnopall)! ! an)! (3 = l' ' n) (9)
and
;fziakolf
&. = _' (15)
.vhere
le: 0 ' fr,: 0
or from (18)
" x
fnnaat > (21)
for not all x /ak equal to zero ln words thls compound term
consnstmg of the welghted sum of our unknown rates of change
must demtely be posntne m Sign However thts does not add
very much to our knowledge smce we do not know whnch terms
w1ll be posmve
As mentioned earlier however we are not always Interested ln
parameters wh1chShift each and every one of our ethbnum con
drtxons For thxs would necessrtate a knowledge of the relatne
quantxtatwe importance of each shlft before we could hope to
evaluate the composuze result For thts reason we often mrrow
our problem by consxdermg parameters Wthh cause only one of
our equnhbrtum equatlons to shnlt Let us restnct our attentlon
therefore to a set of such parameters (a; a,.) equal m number
to our unknowns We may number each 111the order correspond
mg to the respectlve equlllbrmm equatlon Wthl! 1t shlfts Then
smce a change unthe kth parameter must leaxe all other equatrons
unchanged we have
fx &= 0 for J ?! k (22)
Now our mequahty of (21) reduces to the more Slmple and more
readxly apphcable condmon
ax
f,... >0 (23)
In words for the class of parameters now under conSIderattonthe
rate of change of the kth vanable thh respect to ltScorresponding
parameter must be of the same sxgn as fml \VlllCll
m turn 13posxtwe
nf the Shift of the ethbnum equatlon 15m the dlrectxonof an
merease m x,,
Thls can be vened by the computatlon mdxcated (15)
Accordmg to our present hypothes1s
1332= _fz.a,Hu (24)
am. }!
THEORY OF MAXIMIZING BEHAVIOR 33
In Mathematical Appendix A, Section III, it is shown that
11k]:
<0, (k=1,-'-,n) (25)
as a condition for a true maximum. Hence,
}: __ HH:
era, 6a);" (fzkak)2
(T ), (26)
or
x,
f, 67; > 0. (2, )
Let us examine more closely the nature of our hypothesis em
bodied in (22) that each parameter shifts but one condition of
equilibrium, leaving all others unchanged. In the rst place this
does not mean that a change in the ith parameter results in a
change in the ith variable alone. On the contrary, a change in
any parameter will typically result in a change in all variables.
Our hypothesis merely says that this must come about through a
shift in but one schedule with movements along the remaining
schedules.
As I shall later argue, the assumption of this hypothesis does
not involve any serious loss of generality and still includes the vast
majority (in fact, it is hard to nd exceptions) of relationships
contained in current economic theory.
The most general function for which the partial differential
equations of (22) hold may be written
Z = 0(x11' ' ' ! xn) + Bl(xly al) + B2(x2) 0:2)+ ' ' ' + Bn(xm an)' (28)
Ther dz " dx a
36::sz d"; +f.==0+fa=a: (32)
That 15 the rst order change m 2 ISexactly equal to the change
z when the x 8 are not varymg optimally so as to keep z at a
maXImum only to terms of a higher order 15there a dlfference 111
the way 2 as changing A snmlar relanon can be shown to hold
m the case of a constralned maxnnum
This 15the famlhar relatxon of tangency between the envelope
of a family of curves and the curves whzch1ttouches It has many
economlc applicatlons of whtch only a few need be mentioned In
the famous dlspute between Professor Vmer and lns draftsman
Dr Wong the question arose as to the correct relatlonshnpbetween
long and short run curves The short run curve ISdrawn up by
mtmmrzmg total (and average) costs for each and every output a:
of gnen amounts of somexed factor The long run curve requires
lowest total (and umt) costs for every output as plant xsoptxmay
adJusted But accordmg to our theorem a change m the parameter
output wrll to the rst order result m the same change total (and
mm) mais mrerr pn'anrrs
xed aswhen pxcmtas @ma'djaw
Dr Wong was right then msnstmg upon tangency
Another example rs provxded by Professor V1ner' Let labor
marginal cost be the rate of Increase of total costs when only labor
' ] V ner Cost Curves and Supply Curves Ze Lcchffur Nalxmalokonom III
(1932) 2346
" ] V ner Stud :: m the Theory of In amat ouai Trade (New York Harper 1937)
516
THEORY OF MAXIMIZING BEHA VIOR 35
is varying so as to produce the extra output. It is to be dis
tinguished from marginal cost of labor which appears in monopsony
theory, and the marginal labor cost which means the increase in
one component of costs as all factors are Optimally varying. Let
marginal cost be the rate of increase of total costs with respect to
output as all factors vary optimally. Then our theorem states
that labor-marginal cost equals marginal cost equals any other
factor-marginal cost. As Professor Viner has pointed out with
great insight, at the margin (i.e., neglecting differential coefcients
of higher-orders) all factors are perfectly indifferent substitutes.
The classical economists, lacking the precise notion of an inni
tesimal, were forced to employ the concept of a broad extensive
margin. (Viz., Ricardos no-rent land, and J. B. Clarks famous
zone of indifference.)
For nite movements, however small, higher-order terms will
make the change in the quantity maximized (minimized) different
when the unknowns are optimally adjusted from the change when
they are held constant. Actually,
d2z n d2xi n dx.- d(fi) n dx,
E5= Zfi+
1 a lg--l' ailing-Ff (33)
dx,
0 + O+ (if-{a- +fai
since f.; = O. .
The higher-order change in 2 when all x's are held constant lS
given by
2
%=f- @@
The difference between the former change and the latter is as
follows:
m$Mm>
dzz
2z " d_x,-
()35
because of equation (21) above. . _.
If the changing parameter affects prots or ordinal utility ad
versely, it does so less when output and consumption are optimally
adjusted to the new circumstances. If the parameter improves
uFor still other examples see Marshalls Principles, Mathematical Apper
Note XIV, pp. 846-852.
30 rUUNv/U'IUNS OF ECONOMIC ANALYSIS
where
_ [gl]
ISof rank r.
The introduction of constraints requires that our equ1hbrium
system be modied to assume the following form
0i+r)\ f8- i=01 ('l:=l,",7l)
?, BE a (40)
rH= % gil.
0 (B7=1,.,7) (41)
Then the change in x.-with respect to 0:,-when r auxiliary con
straints are imposed, is given by
dx,- _ LH; (42)
(aa.)' % <
if [H] is the matrix of a negative denite form, and the matrlx
38 FOUNDATIONS OF ECONOMIC ANALYSIS
(a-)
da. ' - (a)
da; ,-1 a
'II IH
"FI 'HiH-r. a+! "
'H 'Hu-i-r.n+r
" 'HH'H'N'r' "+' _ 'H'Hn4-r n+ru (43)
_ 'HrH' '. +v
(' )
= __'_1'J'_'._
'H'H'H". 84?
@) (%)
da. o = da; 1 r-
= 4%)
: dat ul
IM O (44)
It is not hard to point out the relevance of the above for a large
number Of economic problems. First, let us repeat the result of
our previous analysis.
6x,
ft.-a.- aai
> 0. (45)
This states that the direction of the change Of the ith variable
with respect to its corresponding parameter is Ofthe same sign as
f,_.._.(
= B). This quantity may be taken as a criterion. If its
algebraic sign is denite, i.e., unambiguously determinable, then
the sign Ofaxi/aai is also denite.
We have merely to show, therefore, that a wide variety of eco
nomic problems can be so formulated as to yield a conclusive de
termination Ofthe sign of our criterion.
It may be noted that for maximum problems involving a single
variable no restrictions need be placed upon 2 in order that our
criterion be applicable. Let
2 = f(x, a). (46)
Then our criterion becomes
x
f" 5&-> 0. (47)
I list below some random examples chosen to illustrate the
applicability of the criterion.
(a) Let us go back to the example Ofthe second chapter Ofthe
eect of a' unit tax on output. It will be remembered that the
relationship between the equilibrium output and the tax was de
termined to be unambiguously negative. This conclusion can be
quickly Obtained by our present method. Prot is dened as
follows:
1r = an(x,t) = Exx) - C(96)] tx. (48)
Our criterion may be easily computed.
W:: = %: = " I. (49)
40 FOUNDATIONS OF ECONOMIC ANALLSIS
Therefore,
6x
( I) 5 > 0 (sa)
or
x
a <0 (51)
Also. as shox the prcvzous section,
dn' 311'
Obvrously. therefore, for the rst case of a tax on gross sales, the
effect of an Increase m the tax rate ISto reduce output The last
nm cases however, present a new feature Our cntenon ISneither
posrtxve nor negatnc, but equal to zero A little reectxonreveals
that our equ111br1ummaxtmum conditions are essentrally unde
pendent of the parameters changed Hence, our equlhbrrum
\alucs remain unchanged, 1e ,
; a 0.
al (54)
x E 0
dl H
THEORY OF MAXHHZING BEHAVIOR 4I
These conclusions are, of course, familiar from the Marshallian
analysis.
(c) Let us now consider a problem which has nothing to do with
taxation, but which gured prominently in the famous cost con
troversy of a few years back. Let us suppose a rm in pure com
petition, i.e., one which can sell as much of its output as it wishes
without affecting the price. Given total cost in function of output,
there will be a determinable output reaction to each given price.
What is the nature of this dependence? From the fact that the
rm is assumed to be in stable equilibrium when it enjoys a proper
relative maximum, we can easily apply our criterion to deduce the
properties of the supply curve. Here
11'= 99(xtp) = px _ C(x)r (55)
1r , _
=p_c@y41 o
Equation (56) can be solved to determine output in function of
price.
x = g(1>) (57)
It is easily veried that
772p E 9911:E 1- (58)
Therefore, d
3% = g(p) > 0. (59)
Even if we relax the requirement that a regular relative maximum
be realized, it still remains true that the supply curve cannot be
negatively inclined. Of course, this does not mean that the mar
ginal cost curve cannot be negatively inclined, but merely that in
such ranges it cannot serve as a supply curve.
(d) It must not be thought, however, that the assumption of
our equilibrium as the solution of a maximum problem is the open
sesame" to the successful unambiguous determination of all possible
questions which we may ask. For it is extremely easy to specify
simple and important problems which cannot be answered even
qualitatively without further knowledge.
Consider the problem of the effect of the introduction of adver
tising expense upon the output of a monopolist. Will an increased
advertising expense result in a larger or smaller output? Here
w=mw=mecmm mm
THEORY OF MAXIJIIIZING BEHAVIOR 43
Under discriminating monopoly all prices are regarded as inde
pendently variable and are adjusted so as to maximize prot.
Prot may be written as
rr= P1D1(1J1)+ P202(P2) "' CEDI + D2(P2)]- (65)
For simplicity we rule out the possibility that all of the demand
at a given price in a given market may not be satised by the
entrepreneur. The removal of this restriction can be easily made.
Here the conditions of equilibrium are
611' , le
=O=DI(P1)+(P1C)v (66)
in; _ _ , dD2
am 0 web:) + (p2 6);;
This results in an optimal set of prices (p10,pg") and quantites
(x10,x2"), and total quantity (X), or (x1+ x20).
In the case of simple monopoly we impose the condition upon
our problem that the price be equal in both markets. Thus,
pl : 12= P: (67)
= 1201(1))+ we) CEM) + Dem]. (68)
where w,-and a,-are the respective prices and quantities of the z'th
factor of production. But we are also given a production function
(assumed continuous),
___.x(al . . ., a). (80)
Hence,
7r= f(a,,---,a,,,'w1, ---,w,.)
= R[x(a,, ---, a,,)] -('wlal + + ma). (81)
Recall from (21) our generalized criterion.
z f" " 60
aw.
>0. (82)
46 FOUNDATIONS OF ECONOMIC ANALYSIS
Obwously here
fm E 0 (z # J) (83)
Hence,
;
f%:>0 (2:10 on) (84)
But
fan. E ' 1 (z = lv . n) (85)
So
aa.
m<0 (z=1 .n) (86)
This conclusmn holds for any number of factors
The above examples are but a small sample of economlcprob
lems whrch can be regarded as determmed by the solution of
max1ma problems For these, the cntena outlmed above may be
employed to deduce meaningful unamblguous quahtatwe theorems
ANALYSIS OF Fmrra CHANGES
Up untll now the analysm has been conned almost excluswely
to the determination of the algebralc sngnsof Instantaneous rates
of change One cannot leave the matter here, for m the world of
real phenomena all changes are necessarlly mte and Instantaneous
rates of change remam only llmntmgabstractlons It 15Imperatwe
therefore that we develop the implicattons of our analysrs {ormte
changes Fortunately, desplte the Impressroncurrent among many
economists that the calculus can only be applied to mmtesxmal
movements th:s ISeasy done
Let us conSIder for mmplncntya functional relatronsh1p between
one vanable, x, and one parameter, a contmuous and twncedr'er
entlable everywhere on a given Interval
x=g(a) aab (87)
Suppose that we have already ascertamed the fact that the alge
branc srgn of the Instantaneous rate of change of thrs function xs
everywhere negatlve on the dened Interval, 1e ,
I: L !
Smcc we shall be consndermg cluefly the selectxon of regul'ar
maxzmum posmons, we may assume
(- 1)"H = (-1)"lfu,| > 0 (99)
at the pomt of ethbrzum Provndmg thls expressnonremains
posxtzveeverywhere we may be sure of a umque equnhbnum Of
course thxs xssufCIent but not necessary
' See any \dvanccd Calculus
THEORY OF MAXIMIZING BEHA VIOR 49
Suppose, however, that the Hessian does change sign a nite
number of times within the region of economically admissible
values. Then the functions in (97) will be multiple-valued with a
nite number of branches. Some of these may be ruled out im
mediately as not constituting maximum positions, namely those
for which
( 1)"H < O. (100)
It is possible to choose from the remaining branches only by re
ferring back to our original maximum problem. Let (x11,- - -, x,.l),
(x12, - -, x,.2), - - - correspond to a preassigned set (al, - - -, a) as
multiple solutions of (96). That set (or sets) is retained for which
f is greatest. This will ordinarily serve to dene our xs as single
valued functions of the a's except at a nite number of points.
At these points it is a matter of indifference which of the possible
alternative solutions is embraced.
Broadly speaking this possibility of multiple equilibrium offers
no serious difculty. For all qualitative results remain. To illus
trate this let us consider the problem in full generality. Let
z = f(xx, -,x
on. an) = 0(x1. ---.xn) + B(x1.a1)
+ 32m, a2) + + B(xan)- (101)
Considera preassigned set of values of our parameters (021,- - -, an),
and a corresponding optimal set of values for our unknowns
(x10,", xn), not assumed to be necessarily unique. Then by
denition of a maximum
j(xl! ' " xny alu, ' ' ', duo) ; f(xloy ' ' ', xnopalor ' ' '! ant!) (102)
L: EJ fz,(xh ' ' " xl" al' ' ': an)dxidaj 0- (107)
But from (101)
fac; = Be"; 5 or (i # j)
rm = B 8".' (108)
0' J
? f.:, f .* fx,c(xlv' ' : x!"ab ' " al)dxodai 0. (109)
By the theorem of the mean this may be written
! Kauai a o. (no)
where : is evaluated at an intermediate point. If we consider
movements of the kth parameter alone, others being held constant.
this becomes
Kama, o. (k = 1, --,n) (111)
Hence, if our criterion is denite in sign everywhere, e.g.,
fa'b : BHO: < bn (k = It ' ". 13) (112)
then,
Anga. g 0. (k = I, - - -, n) (113)
This is proof that multiplicity of equilibrium values does not
alter the deniteness of our conclusions with respect to nite
changes. It also shows that the criterion which we applied m
earlier sections to determine the deniteness of instantaneous rates
oi change can be generalized to determine the deniteness of nite
changes. In fact, by a proper limiting process our prevmus
theorems relating to instantaneous rates of change can be deduced
as special cases from this more general analysis. _
The equality sign can be dropped il Aa # 0, and if the-main
mum is proper. It will be noted that in this proof we d1d not
THEORY OF MAXIMIZING BEHAVIOR 51
require that 0 be continuous, but merely that the B be continuous
with the derivatives of the required order. This will turn out to
be of considerable importance in the later consideration of discon
tinuities in the production function of a rm.
A very important case is that in which 2 takes the form
AaiAx; ; o. (118)
For this analysis it is not necessary that the cost curve be con
tinuous, or be such as to yield a unique optimum output. The
marginal cost curve may be undened at points. have kinks, and
turn up and down many times.
ANALYTIC FUNCTIONS
(>
<->
(i = 1. -- .n) (123)
The general coefcient is oi the form
( dal! "
. . . dan!- )
where
}: m. = n. (124)
l
" For example. U) = 0. (i = 1, ' ' '. 7!)or FU) = 0, where F(0) : O; F(a) # 0'
a # 0.
54 I'OUNDATIONS OF ECONOMIC ANALYSIS
ku = % (132)
Also for every possible triplet
). l\ )i . = f ufcf 1:
= " fk f k"
, t * fnofuofkko fuufk'co n = Ankh-[Am (133)
dF = 0 implies among other things that f(x.. , x,.) = O This is denitely not an
extremism pasition as shown by reference to the secondary conditions The fact that
in a larger set of variables a stationary value corresponds to a non specmlizedsystem
seems devmd of economic Signicance See G D BithoH, Dynamical Systems (New
York 1927). Pp 3334
THEORY OF MAXIMIZING BEHA VIOR 55
As we have seen, it is not enough that our equilibrium equations
be expressible as the partial derivatives of some function. In order
to be able to derive denite theorems it is desirable that this func
tion be at a regular maximum or minimum. This requires as is
shown in Mathematical Appendix A, Section II, that certain
quadratic forms be either negative or positive denite. Thus, for
a maximum
for 13,j, and I: all unequal. This proves that it is not important
that our problem be either a maximum or minimum one, but only
that it be one or the other.
From (138)
aR R'
ap, " ap. )
There exists, therefore. a function
Z = RG1! ' : pu) .- (V1P1+ Vng + ' + VRP-). (141)
where
3613
?. = Rm. m. (.- = 1. n) (I42)
The conditions of equilibrium represented by the general demand
functions of (139) are, therefore, equivalent to those derived {tom
the condition that Z be at a maximum with respect to the 9s, 1.e.,
i ; (8)
67.0,- 6w,- ' x
CONDITIONS or EQUILIBRIUM
Thus far we have employed only the notation of mathematics.
By so doing, the problem has been clearly framed, which in many
economic problems is more than half the battle. It remains now
to state the derivations of (4) and (5) from (1) and (3).
Our problem is to minimize
C=A+iww. (9)
l
subject to
@(01, ' , v..) = a:= constant. (10)
Mathematically, this is a constrained minimum problem, and we
may avail ourselves of the method of the Lagrangean (undeter
mined) multiplier. We dene a new function
Zx Z1 (0115351< 0! (14)
for X
.. L/
? PIE: = ou
and not all
! : 0
Con51derthe bordered determinant
9011 Pu (Din 991
35" 52. 21 9922 Wu 90:
= (O.; P. =
D l P: 0 (15)
051 ion? Pn Pn
so: 992 p.. 0
, gaJdv, = dx.
1
where
(Pu 9912 99m
Pz
ga-l
I, 99- _- (pal
o 9022
. (.021:
. )
.
A : gai _- _ _ . . (20)
O .
and A.Iris the cofactor of the element in the rth column and the
qth row.
64 FOUNDATIONS OF ECONOMIC ANALYSIS
Also,
" dw,
dh = zx:T A +1+ dard, n+1
(21)
Hence,
'._ &
w, _ XA (22)
As a specnal case
59.: _ ka
awk ... A (23)
Also,
% Ami1k
69: A (24)
and
_a_A__ A} n+1
awk _ RA ' (25)
Q = Amt-ln+l
x A (26)
That is, any xed output will always be produced with less of any
given factor as its respective price rises, other prices not changing.
By the law of the mean this can be shown to hold for nite changes.
Let us now determine the economic meaning of A. Rewrit
ing (11)
G = A + fwd); _ ESOJI,
"'!vn) ' 5]:
1
dC = w,dv,-, (35)
1
and
dx = i tom. (36)
].
Mada),- .
C __l____ = Aer, w!) . . ., w"). (
ax _ tpzdvi
l
66 FOUNDATIONS OF ECOAOMIC ANALISIS
More rigorously the proof xsas follows
EE .. 93
6.1. _ . to, 6x (39)
Substituting from (24),
_ Art-F1I _ * An+1 .
BJC21:10. A >\Zlcp. A (40)
But expanding A m terms of the elements of the last row,
Hence, C A
& = XE = X (42)
Therefore, we may rewnte (12) as follows
6C
20. = a Pt (43)
m; : exam
From (24), (25), and (29),
vk
_wg" ax (46)
or
2
C 02": (47)
exam= 5
Thxs has been pomted out In another connecnon m lectures by Professor Vmer
wnh rllummatmg mslght mto the relatronsth of the external to the Internal margm and
the broad zone of Indlflerenceas a subsntnte l'or the mmteslmal Pendant-ally, It u
du: condition whtch ISbeen: to Mr Wong : famous envelope theorem]
THEORY OF COST AND PRODUCTION 67
That is, the change in any input item with respect to an increase
in output must be equal to the change in marginal cost with respect
to a change in the price of that input item.
Recall from (26)
ax _ 62C A
ax_ a? T ' (48)
Now it is known from (17) and (27) that A has the sign of ( 1)".
Also
sen sa
Am. = - - = H, (49)
sent sa.
_av,- < . _!
v __
av], _ ___?
69- 2 .
a'wj , wj awk (awk) > 0' etc. (60)
It may be well to summarize our results of this section:
621- a i!
4<
wf 0; M)
6(w1') wk) >0; etc. (34)and(59)
ave _ %
aw,- _ wk (31)
62 _ al,
xw; _ x ' (47)
6C .
w.- = Mp.- = 3; a.-, (z = 1, -- -, n) (12) and (42)
FC
( 1)H-x2< 0, (50)
" v; .
jwj=0, (z=1,--o,n) (53)
C
&: = I" (55)
" 6C
C=A+lewrawi' (54)
Even in the case where the production function and its deriva
tives are continuous with the proper convexity to insure a uniquely
determined optimal position, the interesting case may arise where
some factor may not be used at all. That is to say, the more that
others are used and the less of it, the lower will costs be for any
given output. In this case the conditions for equilibrium do not
require the equalization of the marginal productivity of the last
dollar spent on it to that of other factors. Rather do we have a
boundary minimum due to the fact that no negative values are
economically admissible. Hence, the conditions of equilibrium
are given by the statement that for any input, potentially usable
but not actually used, the marginal productivity of the last dollar
7O FOUNDATIONS OF ECONOMIC ANALYSIS
where (vf, .. -, 12.2)is any other point along the same isoquant.
This merely says that there exist one or more tangent Planesat
each point which touch but never cross the isoquant. Similarly,
at the second point there exist constants, (af, . -, au:). sud! that
01' ;
Since these are two arbitrary points, taking one as xed, say
(vf, - - ,Uno),we must have along an isoquant
a,Au. :$ O (71)
l
{or
Olo! . ' ': vo) : 500,10+ Av! ' ' ': au + Av"):
THEORY OF COST AND PRODUCTION 73
and
>:
1
AAU. 0. (72)
l
conditions imposed on (a1, ---, an) in order that (71) holds are
the following inequalities:
(Pis S Oli SoiL
IM , (73)
SOIL 010 %s
CONDITIONS OF EQUILIBRIUM
Suppose we are given a set of prices (101,' ' ', 'wn")correspond
ing to which there is one combination of factors (11?,- - -, v,.)which
minimizes total cost for given output. As a denition of our
minimum
AC 0 (75)
for
Ax = 0, Av.-g 0.
In other words, for any point (vl, --, v,.) along the same isoquant
we must have
for
(vlg . . ., vn) : P(v10, . . ., vnO)
or
tv,-"Av; 0
l
74 FOUNDATIONS or ECONOMIC ANALYSIS
for
4910+ Avg, - . -. u." + Av..) == (vi. ---. vn)
Obviously from (73) this implies
geg<ei
8
mm. 2 0 (82)
1
for
Ax = 0.
Suppose that only the price of the lath factor varies. Then
(82) becomes
4103A; 0. (83)
That is, an increase in the price of a factor cannot result in an
increase in its use. Likewise, a decrease in its price cannot result
in a decrease in its use. Still more generally, it can be stated that
a change in the price of any number of factors cannot result in a
change in amounts of all the factors in the same direction. i.e.,
%; = 52%
g o (88)
The results m (96) follow from (43) and (89) smce for any
output factor price must equal margmal cost tlmes margmalphys
real productmty whrle for Optimum output marginal cost equals
marginal revenue These are n equations from whrch we can
solve for the n factors of production 1nterms of the n puces to get
the demand functions of (94)
I). ::= g'(w1 . un) (Z= I, , )
Actually, 1t lS known that "
_6v. I2
6w, = gli : T : (97)
where T., 15the cofactor of the element of the zth row and the 3th
column of the above mama From the denntenessof the above
quadratrc form It follows that these last must form the coefcnts
of a negatrve demte form, 1e ,
g; 6(g. g")
aw, < 0, (TH(w,20;) > 0, etc (98)
_ A _ ww. (103)
l
THEORY OF COST AND PRODUCTION 81
Now adding (102) and (103) and canceling terms we get
ZZjAwiAv;g 0. (105)
For only the jth price varying, this becomes
Aug-Av;g 0. (106)
In other words, a decrease in a price cannot result in a decrease in
the factor used. Further possible interpretations are of course
possible.
As before, the general case is simpler than that of the special
continuous case. Moreover, the method of nite increments ap
pears to be mathematically simpler in the sense that it is possible
to state the qualitative direction of changes without solving in
versely for the actual demand functions.
The method employed here is that which underlies Le Chateliers
principle in physics. By making use of Professor E. B. Wilsons
suggestion that this is essentially a mathematical theorem ap
plicable to economics, it has been possible to gain increased gener
ality without increased complexity and emptiness.
It is important to realize just how much content there is to a
particular economic theory. As far as the single rm is concerned,
everything fundamental which can be said is implied in the state
ment that in equilibrium there must exist no movement by which
the rm can improve its prots; i.e., A7:- 0 for all movements Of
variables possible to the rm. In the case of continuity certain
necessary relations of differential coefcients (marginal equiva
lences) are implied. Moreover, assuming certain specic forms to
our functions (independence of prices, etc.), it is possible to deduce
formally the implications of an equilibrium position (e.g., negative
demand curves, positive supply curves, etc.). It appears that no
more than this can be validly stated.
EXTERNAL CONDITIONS or EQUILIBRIUM
Thus far we have been discussing the conditions of equilibrium
imposed from within the enterprise by its deSIre to maXImize
82 FOUNDATIONS OF ECONOMIC ANALYSIS
prots Thrs has resulted m certam margmal mequahtles Econ
omrsts have not stopped here, but have also tried to analyzecertam
condrtrons of equnllbnum resultmg from mtercompetrtxon among
rms That 15 to say, they have tried to state COHdIthHSupon
the market sxtuatxon (obstacles) thh WhiCheach rm M be con
fronted In particular, they have been mterested 1nthe determlm
tron of the rate of prot which any rm can earn
It has often been argued that not only must price (average
revenue) under perfect" competxtlon equal margmal cost. but also
It must be equal to average cost so that net revenue wnllbe zero
Thls second condrtron has not always been recognlzed as beingof
an entrrely different nature from the rst In tlus sectlon an
attempt wnll be made to dlStIHgUIShbetween them It IShoped
that m so domg xt Willbe possrble to put the famous addmg up"
problem and homogeneity of the productron functron m ntsproper
place
In the beginning, to avond confusmn no use 15made of the
term perfect ' competltron The term pure" competltronwrllbe
understood to mean that the demand curve for any producer 15
1nn1tely elastlc that his sales cannot affect prnces The problem
of drscontmurty 1s 1gnored Under these condttlons the Internal
condltxons of equlhbnum are that marginal cost be equal to mar
gmal revenue (pnce) and hence that the marginal physrcalproduc
trvrty of each factor tlmes the sales price of the good be equal to
the pnce of the respective factor
: : lg , (107)
ax x
C
w. = 5 P= Pv. (103)
These are margmal condltlons and say nothmg about the totals
'mvdrw. Lt.'sden. tat-m.h}; demrmn of. the long; run" as that
perrod m which all costs can be avorded by goxng out of busmcss,
that the rm must never have a negatwe net revenue As a
condmon mtemally Imposed we know that
fr 0, (109)
or
R(x) ; uw. (110)
l
THEORY OF COST AND PRODUCTION 83
Some writers, by a curious play on words, have been able to
arrive at the condition that average cost equals price. A typical
form of the argument is as follows: (1) a rm will equate marginal
cost to price; (2) it will also try to minimize its unit cost; (3) at the
point of minimum unit cost average cost equals marginal cost;
(4) hence, average cost must equal price (average revenue) and
prots will be zero.
Stated explicitly, it is obvious that the second statement is false.
The play on words arises from the confusion of the condition that
for each and every output total and unit costs must be a minimum
with the statement that of all outputs possible that one is chosen
at which unit cost is the lowest. The rst implies the valid condi
tion that the marginal productivity of the last dollar in every use
must be equal. The second implies the invalid (from internal con
siderations) condition that output be determined irrespectiveof the
sales price.
On the other hand, some have tried to argue as follows: (1) the
production function from the nature of things must be homogene
ous of the rst order; (2) by Euler's theorem it follows thatf if
factors are paid according to the marginal productivity principle,
product will be exhausted. .,
As an example of the lack of integration between the theory of
production and that of cost, we nd many writers asking whether
product will be exhausted at the same time that they have already
agreed that price equals average cost, and total revenue equals
total cost. Of course, the latter condition is merely another way
of stating the former.
Once the problem is properly stated as that of determining the
relation between gross revenue per unit and expenditure per unit,
it should be reasonably clear that this cannot be determined by
the properties of the production function alone, but must depend
upon the marketing situation of the rm, which in turn depends
upon the competition of other rms. It is quite clear that as far
as the single rm is concerned it is possible that it be making huge
prots regardless of the homogeneity of the production function.
This condition is neither necessary nor sufcient to the exhaustion
of the product. If the production function were homogeneous.bUt
demand were sufciently favorable, of course product wel
be exhaustedeven under pure competition.
84 FOUNDATIONS OF ECONOMIC ANALYSIS
minimum average cost. This follows from the fact that average
cost must be equal to marginal cost, and that the latter is rising.
6R 6C C . .
5; = p = ?; = ; = minimumaveragecost." (113)
0f course,
x = ? (pv. (116)
x M., (117)
wM:
Ol' U
Barge- ___:= x 25)
P1 P2 _ p:- (
Th1s means that m cqunhbnum the ratlo of the marginal utnhttes
of two goods IS equal to the ratxo of thenr pnces, 1e. margmal
utlhtles are proportxonal to prices
It 15clear from the formulation of (24) that lt does not matter
whtch utxhty mdex we use, for
U: = FRO! (! =1v ' ; n) (26)
THEOR Y OF CONSUJlIER'S BEHA VIOR 99
Therefore,
of (19) must
librium can bealso be Satised. and so our full conditions of equi
written
U.,dx,-
+ pm = ( mph (2'= 1, n)
1 (32)
where
Un Un Um P1
U21 U22 U2n P:
D _ _U_l
Pi 0!).-| , I. . . . (34)
Unl Un'z ' ' Una Pn
P1 p.) ... p 0
and Da-indicates the cofactor of the element of the z'th row and the
jth column.
102 FOUNDATIONS OF ECONOMIC ANALYSIS
Kit = 3%+ xi % ?
It has been called by Slutsky the residual variability of thejth good
for a compensated change in the ith price.B
This can be made more clear by the following considerations.
Thus far, we have imagined the individual to be maximizing his
utility as of given prices and total expenditure. A little thought
will reveal that utility will be maximized as of a given expenditure
only if the level of utility which is being realized is being achieved
in the cheapest possible way; i.e., expenditure must be minimized
as of any level of utility. If this were not so, the same level could
be achieved with some money left over; this remnant could be
spent to buy more goods, and hence a still higher level of utility
could be attained.
Along any indifference _locusthere exists for any set of prices
an optimum set of purchases which minimizes total expenditure.
That is,
x: = W'Dbi.up. F(so)]. (J' = 1, "un) (43)
For
U = F(o) = constant,
we are conned to the same level of utility. It could be easily
shown that
Ich-=(91)
P. U=constant sw: (i.j=1.---.n) (44)
In words, K5.-is equal to the change in the quantity of the jth
good with respect to the z'th price, where the 1nd1v1dualmoves
' E. Slutsky, Sulla teoria. del bilancio del consumatore," Giamale degli economisli,
LI (1915), 1923.
See chap. iv, pp. 62-65.
104 FOUNDATIONS 0F ECONOMIC ANALYSIS
along the same indifference locus and keeps lus expenditure down
to a mmxmum before and after the change m pnce "'
From the formulatzon of our cqwhbnum equations un (29) we
have seen that the properties of the demand functzons are nal
accted by our choxcc of utlllty Index Thus may be shown et
phcutly from the Identttxes
U=H @@
U: : F'. (46)
Un = [3500 + Fn'P191 (47)
( A) = F'( A'), (48)
where (--A') xs the marginal utlhty of Income for the unhty
Index a Let
DI s: lLli
Pi 0pl, (49)
From (47)
D={1UJ
P: {={Ml
0 m (Al:Peal
0 p; P.!
0
s (Pr'D'
Snmxlarly the following relatnonshnp holds for all cofactors
Do = (JD.; (1 J = 1. . ") (51)
Hence
(- un ___(mon, (52)
D D
so that K,. 13 an luxuriant under any transfer-manon ol utxllty
Index Lemony nothmg :: :mphed jar cmpmca! pme behamr 5?
the chanceof any partmdar uhhty index
Inspection of our ethbrmm condmons m the form of (24)
and (19) reveals that they are unaffected by a proportional change
In all puces and mcomc our equlhbnum values remam Intact for
such '1 change le .
x.=h(p. p.. I)=h'(mp. mp. ml) (:=l, .it) (53)
here m as any posxtwc number Mathematically our demand
For another lntcrprcmt on see H Schultz The The"? ! anrcmml of
Demand (Ch (ago Univers ty of Chicago Press 1933) PP 4345
THEOR I' OF CONSUIER'S BEHA VIOR 105
functions must be homogeneous of order zero. Employing Euler's
theorem for homogeneous functions, we have "
6x,- _x_.- x
a__. ax _
:>-+--+
6911"+p-;b a+
I= O. (z=1,---,n) (S4)
Dividing through by x.-yields the following relationships in terms
of elasticity coefcients:
17i1+m2+''+77in+77u=0. (i=1,---,n) (55)
where
_ % Pi
Tm _ 6171'x.
H(I'l
_ 3 &: I&) (1= 1, ,n)()59
From our def nmon of meome or total expenditure as
I - ZM
:
we have the {allowtng(n + l) restrictions on ehsucmes of demand
Z I! : = 1
l
and
.
: &u = - h
I
where
x
7
l
13the proportion of mcome spent On the :th good However these are not mailing;
restnctwns state they are consequences of our denmon At best they C
reveal that we have not applied our dened operations With numenall accuracy
THEORY OF CONSUMERS BEHA VIOR 107
Let
a,
u
NI?
Q n r" : V
Hence,
xi = Hlah ', a"). (60)
The as here are very natural units to employ, since they involve
only the dimensions of the respective quantities. In words, a,-may
be dened as the proportion of total income required to purchase
a single unit of the jth good.
II. We have also the following reciprocal integrability con
ditions.
x- x- ax.- ax.- . .
Kj'=P+xii=;+xi=Ki (11.7=11"'1n) (61)
i.e., the residual variability of the jth good for a compensated
change in the ith price is identically equal to the corresponding
term for the ith good with respect to the jth price. These are
n(n 1)/2 independent meaningful conditions.13
MEANINGFUL THEOREMS
Thus far almost nothing has been said about the directionsof
change in our equilibrium quantities of goods demanded with re
spect to changes in prices and income. Does the utility anaIySis
have nothing to say upon this question? The answer can be sought
along lines indicated in previous chapters.
Before clouding the air with determinants, let us make a
common sense appraisal of the situation to see whether we cannot
suggest a simple answer.
First, suppose the individual to be constrained to move along
the same indifference locus. Let him be confronted With a set of
prices, and attempt to attain this level of utility" in the cheapest
They hold, of course, only for the individual demand functions. Moreover, they
reect differential properties of our demand functions which are .hard to Visualizeand
hard to refute. For our empirical data consists of isolated pomts. _The must be
smoothed in some sense before our relations an be tested; the smoothing. even by the
best known statistical methods, is to a degree arbitrary, and so refutation and verication
are difficult. _ _ . ..
I have tried, but thus far with no success, to deduce implications of our mtegmblllt
Conditions which can be expressed in nite form, i.e., be conceivably refutable merely
by a nite number of point observations.
108 FOUNDATIONS OF ECONOMIC ANALYSIS
possrble way Consrder the set of puces (91, , p..) There
wrll correspond to this an optimal set of quantltles (n, ,xg)
such that total expendrture rs as low as possrble : e ,
In equatlons (62) and (64) any values of x (along the same locus)
may be Inserted the respective left-hand srdes In particular,
we may write them respectwely
m.! >:pm. (65)
I 1
and
m. 2":px. (66)
I. l
This means that the optimal set of goods for each respectzveset of
pnces cannot cost more than the other set of goods (optimal fora
dlerent set of pnces)
Rewrrtmg the equations, we get
p.(x. x.!) s 0, (67)
1
Suppose we allow but one price to change, say the kth; then
all but one term of (71) vanishes and we have
AdQ-Apt. < 0; (k = 1, ' , 71) (72)
i.e., as the kth price increases, all other prices being held constant, less
will be bought of the kth good. It must be emphasized that this
holds only for a movement along the same indifference locus, i.e.,
for a compensated change in price, and does not mean that with a
givenmoney income a change in one price will necessarily result in a
decreased amount taken of the corresponding commodity. It will
be noted that the above proof does not involve the calculus at all;
usingonly the operations of addition and subtraction, the denition
of a maximum position may be utilized to derive meaningful nite
demand restrictions.
Employing only the most elementary logical and arithmetical
operations, we can advance matters still further. Consider any
initial set of prices and income (1210, ', Pn, I)- Corresponding
to this there will exist one or more optimal sets of goods. Select
one of these and designate it by (x1, ', xn0)' Consider now a
second set of prices and income (1211,- ' ', Pnl, Il), and a correspond
ing Optimal set of goods (x11, - - ,x,.l).
Let us consider what would have been the cost of the second
batch of goods at the prices of the rst. This will be
If this cost is equal to or less than the amount of money that the
rst batch actually cost, we have conclusive evidence that the
second batch is not higher on the individuals preference scale than
the rst batch; for if it were, the individual could not have been
in equilibrium in the rst place, since he would not be mlnimizmg
total expenditure for the attained level of satisfaction. In other
IIO FOUNDATIONS OF ECONOMIC ANALYSIS
i px.< p.x.. (m
and I l
br
f: 9.41:.< 0 impliesif (p. + %. < 0. (82)
! 1
;??st : m px (85)
Also,
N = "2.0. (2 = 1. .n) (86)
Therefore,
But also
$: Nx. = 5i pm (88)
1 1
)::pdx. ==0,
lot all do.. or dp, ==0
TIIEOR Y 0F CONSUMER'S BEIIA VIOR I I3
In this expression the dx's and the dps are differentials, not
innitesimal increments.
Regarding prices and income as independent variables, from
our demand functions of (20) we have
dx.- = >":hjdp;
1
+ ;;de (i = 1, n) (92)
OI
i z (mama.- g 0. (96)
1 1
)5 K.,dp.dp, g 0. (103)
l !
>":2":A.,th, ; o (105)
1 I
" Only a sketch of the prooi ol this statement need be gwen Wme
:. - !!!: val)! ( . le Un)
_ an Ox; ax. ' 95:
A"-3-p-,+357-1(Ba, a:-'n).
Dene a new set of variables
o " f(a|| ! au). or d'. - Filu c1)
such that
" _ ]IEpth ,n). . Pal. IS.)] : Gwh op.)!
and a;
0.1" , _ = _:
65;5 G @ l,! = A! . C! a
Then there exists a inaction
P - (Bl I n)
or
9 " @(xh 01.0)
u hich satises the properties of our preference eld
THEOR Y OF CONSUMER'S BEHA VIOR I I7
where the equality sign holds only for all prices changing in the
same proportion.
Many writers have held the utility analysis to be an integral
and important part of economic theory. Some have even sought
to employ its applicability as a test criterion by which economics
might be separated from the other social sciences. Nevertheless,
I wonder how much economic theory would be changed if either
of the two conditions above were found to be empirically untrue.
I suspect, very little.
F(xl,...xn,.l.,...,E
III M) (113)
where pnces, mcome, and Interest 7' are given to the mdmdual
Thus, the amounts pend out (or foregone) m every perlod for the
use of money are treated as subtractions from Incomeavailable for
expendxture upon consumers goods The condrtlons of equl
lrbrlum are exactly as In chapter v equation (29), except that the
margmal utdmes of the goods are affected by the level of pnccs
dlrectly, and we now have an addrtronal unknown, M, to be de
termmed But we also have an add1t1onalequatton
E _ 115)
6M + xpr 0 (
LU aU. w.
Unit 611! Pu Edi-A6 "gbm 0 0
Pk Pm 0 _; x, r! -1 p,..M
over dram'mc way of saymg that nobody would hold money and
1t would become a free good to go mto the category of shellsand
other things which once served as money We should expect too
that rt would not only pass out of crrculatlon but rt wouldceaseto
be used as a conventional numeralre ln terms of Wthll pnces are
expressed Interest heal-mg money would emerge
Of course the above does not happen m real lnfe precrselybe
cause uncertainty contmgency needs non synchromzatnon of
revenues and outlay transactlon fnctrons etc etc all are W1thus
But the abstract Specralcase analyzed above should warn us agamst
the facrle assumptron that the average levels of the structure of
Interest rates are determined solely or primarily by these dlffer
entra] factors At tlmes they are primary and at other times such
as the twentles thls country they may not be As a generalxza
tron I should hazard the hypothesrs that they are hkely to be of
great xmportance m an economy m whtch there rs a quasr zero
rate of Interest I thmk by this hypothems one can explam many
of the anomahes of the Umted States money market m the thntles
A further defect of the reasoning that I have been crrtrcrzmgles
In the tendency to regard as universal the hypothe51sthat Interest
yreld 15mversely correlated wrth an asset s nearness to money so
that normally long rates are above short rates This 15Of m
'uccord thh much of economic history for reasons whrch I do not
think are difcult to elucndate 5
" See D Durand Bas c Y elds of Corporate BOnds 1900-1942 Techn l Paper 3
016: York Nat onal Bureau of Eoonomc Research 1942) Also F A Lutz The
Structure of Interest Rates Quarterly Journal of Econonuc: LV (1940) 36-63
CHAPTER VI
[it]
aw, = [R.,]I. <3)
Since the inverse of a symmetrical negative definite matrix is itself
Actually it is a little misleading to say that an elasticity expression is necessarily
without dimenSion " For take any absolute derivative,such as dr/dp, which is cer
tainly not dimensionless, involvmg as it does the dimensions [output times output dwidcd
by value] Even though it has dimenSions. it is still the elastiCity of same expression
Thus if
= = KP).
and
y = 5(9).
here
y = "i q =":
then
E_y dr.
Eq'd
TRANSFUKMA 'I'IONS, COMMODITIES, RATIONING I27
symmetrical and negative denite, the complete conditions on the
demand functions are summarized in the above expression.
If one were interested in the corresponding elasticity coefcients
[Evg/ij], one could as an afterthought introducesuitable factors
into the matrix of equation (3). This usual procedure has been
termed by Professor Lange the indirect method.5 He suggests
an alternative direct" method. I should like to point out a third
method, which carries out the Lange procedure to its logical con
clusion. Before doing so, however, I should like to express the
opinion that in this matter indirectness is a virtue rather than
a drawback.
It will simplify the discussion if we adopt a mnemonic notation
whereby the Jacobian of one set of variables with respect to another
is written in a form reminiscent of that for a single derivative; i.e.,
. _ % _ d_3'
J(y11'.')y!xh"'1xn)-[axj:|dx' (4)
The reader may then verify the identities
1342 _ l
dydx _ dx
5
d_x__
dy avr
_ dx ( )
which parallel exactly those for single derivatives.
In terms of this notation the Jacobians of the transformations
Vi = 10 711', 2),- = el'i _
W.- = 10:20.. w.- = e"'-' (* = 1, n) (6)
can be written as
@. 'U
_[2] = [Wk]
dv _ Ui ,
W 5," w
FL%l
% = [217]- aw = Wi
where S,-,-
is the Kronecker delta, equal to one for like subscripts
and vanishing for all others.
Osr Lange, Theoretical Derivation of Elasticities of Demand and Supply: The
Direct Method." Ecanometrica, X (1942), 193214.
128 FOUNDATIONS OF ECONOMIC ANALYSIS
Then It follows that
dv dVdDdu 5: _
m = W= [Elm] 'WJ 8)
Tlns Is the so called 1nd1rcctmethod The nge direct method
mvolvcs computmg the same matrn by the followmgcquahty
dV __ dW 1 v,R., 1
an? [W] = [r] (9)
There seems to be no specxal advantage m thmsprocedure on
the other hand there rs the great d151dvantage of loss of symmetry
prior to mvcrszon and a cloakmg of the demteness properttes of
R and Its Inverse Of course. what IS coxercd up can later be
unscrambled. but a good deal of waste motton would seem to be
Involved
However If e wxsh to carry the direct method to Its nal con
clusmn 1t would seem logtcal to replace the v's and 205 m the
ongmal prot crpressnon by theu equwalences m terms of the
variables V and W \thh the latter being gwen, the rm would
vary the Vs so as to mawmlze prots The condxtxonsof equr
hbrmm would be
S.(V1, V,.) eweI ==0, (z = l, , n) (10)
and
[SU _ v.w..,]: H
negative demte In the above expressxons xt xsto be understood
that
S(V1. . V..) = R(e"'. . e). (11)
and that subscnpts stand as usual for dl'erentxatron
If we dterenttate our equlhbrrum condmons we easdy derive
the 1dent1ty dV
.E'JV
=-..H'[v,w,,,] (12)
H,, = Cf.-...,,
+ {gnaw}. (26)
where the Asare Lagrangean multipliers.
After transformation we have
- ["]
Pi
if _' [
dp l
0 10]
1
_ [ 2l
61_ 21 ] (53)
c get
dx 96 96
?? = o 96] (54)
Much ISneither perfectly stable nor Imperfectly stable
Of course :( [dx/dja] had been symmetncal and demte (as
when Income effects can be neglected) thus could not haxc hap
pened but then our system would have been com ertlble Into a
maximum problem
I have shown else here that convexity non mtcgrablc cases
lmphes that the symmetncal part' " of certain matnces be
denite This follons m consequence of the relation
Value and Capital will take its place in history along with the
classic works of Cournot, VValras,Pareto, and Marshall. Like the
latter, Hicks has succeeded in keeping formidable mathematical
analysis below the surface of things and locked up in appendices,
thereby securing for his work a much wider audience than would
otherwise be possible. This tour de force was made pOSSIble1n
considerable degree by the repeated use of the already mentioned
theorem relating to the demand for a group of commodities when
their prices all change in the same proportion. We are now 1n a
position to derive this theorem rigorously in a more general form.
142 I'OUNDATIONS OF ECONOMIC ANALlSIb
When a group of puces say (1)1 1)) all move together
nt ISnatural to dene a new commodity ::; by the relatxon
$1 = 91361+ P2352+ + P x (59)
:; __ [0PI I[ Pc pI 0 O] x (60)
Here we haxe Simply dened all commodltnes but one to be nden
ncally the same as before the rst commodnty has been replaced
by the new composxtecommodity The contravanant puce van
able must then satlsfy the equatnon
1 O 0 '
-22
Pn
I p (61)
O'-'
0
so as to leave 2 1n: = Z px Even though (n 1) commodltles
are the same as before their pnces must have changed This
wxll not seem strange 11'It IS remembered that a pnce IS not a
property of a good Itself Independently of the frame of reference
mvolved In fact 1t ls qmte possxblefor e1therpr1ces or quantmes
to Lecome negative although ):, pr wdl conserve 1ts ongmal Sign
Now If we change prices (p, p ) m the same prOportlon ).
holdmg constant (p H p,.) It can be shown by exphcxt dr'er
Th 5 s only one of an nn te number of pass ble tmnsformat ons wh ch w ll serve
present purpose
TRANSFORMATIONS, COMMODITIES, RATIONING I43
entiation, d/d), that in the new price-quantity coordinates only
the 131is changed with all other 13sremaining constant. Conse
quently, if this price changes, and if income is changed in such a
way as to keep utility constant, then
:15: _
(TI-97:,
= Ku < 0. (62)
Also
Kquz _ 122> 0, etc., (63)
so that all of the properties of the demand functions given in the
previous chapter are satised.
Thus, the demand or a group whose relative prices are unchanged
satises the same inequalities as in the case of a single good.
This is the fundamental Hicksian theorem.21 It can easily be
broadened so that we may work with s(_<= n) composite goods.
Their substitution terms form a negative denite quadratic form.
This is a very useful theorem. It states that any group of
commodities whose relative prices remain unchanged can be lumped
together into a single commodity, and that indifference curves in
fewer dimensions can be constructed which will have all the usual
properties of indifference curves. Thus, all goods but one can be
lumped into one commodity which Hicks calls money," and then
concave indifference curves can be drawn between the good in
question and money.
It would seem that this involves rather a strained use of the
term money," and one almost certain to lead to confusion.23
Very properly when Hicks later comesto discuss monetary matters,
he rejects this earlier notion.
" Proved in his Mathematical Appendix, pp. 311312,as a consequence of his sixth
rule relating to substitution terms. All six of these rules are contained in the statement
h' [X.-51his non-positive denite quadratic form of rank (n - 1). 'hCh vamshes for
values of h proportional to prices, this being an immediate consequence of the primary
conditions of equilibrium and the secondary conditions guaranteeing an ettremum
under constraint.
Value and Capital, p. 33, passim, diagram on P- 39- 5150see A' G: Hart, Peculi
arities of Indifference Maps Involving Money," Reviewof EconomicStudies, VIII (1941).
126.213258imh
an instance is provided by a writer who interprets Hicks literally and con
{usedly attempts to throw light on the "inflationary gap" by means Ofan indifference
diagram in which money is taken as a mtchall for all but one commodity. M. W. Reder,
Welfare Economics and Rationing." Quarterly Journal of Econonncs, vol. LVII (1942).
144 FOUNDATIONS or ECONOIUCANAL} SIS
Tm: GENERALPnonuzu OF CouposrrE on
AGGREGATECoxmoomas
Any economtc system hen wen ed carefully mll be found to
consrst of an almost uncountably large number of \arnbles It IS
almost a necessrty If ad.ances 1nanalysns are to be made to sumphfy
matters artxctally so as to reduce the number of \arnables whlch
are to be handled T111515done m a number of different mays
Some authors retreat to '1one or mo commodity \\ orld order to
deme more precnseresults the penalty for this hes the difculty
m estabhslung the relation betuecn the sumphed construct and
complex reality Thls ho ever ISat least an honest procedure
Other \xrxters 13hto hate thexr cake and eat It too to work mth
only a few variables and at the same tune retam an "nr of reahsm
and versxmlhtude
Here too a \anet} of artful dodges are open to the mvestngator
(1) He may hold other thmgs equal (2) He may concentrate on a
representatwe rm or family, each of \\ hom IS donng about what
all are domg (3) Or he may \\ ork u 1th certain aggregate, com
posrte magnitudes such as bales of output, "socnall) necessary'
labor age units " cost of hvmg. real natnonal output. etc
lt ns the tlurd deuce whlch 18 of Interest here There :s
nothmg 1ntr1n51callyreprehenszble tn \sorlung \uth such aggregate
concepts On the contrary, abstractron from compleuty xs a
necessary thought process And m any case the most genenl
equlhbnum set up must necessarily stop short of the full total of
all possrble economic \anahles But It 15Important to reahze the
hmxtanons of these aggregates and to analyze the nature of therr
constructlon
Broadly speaking \\ e replace '1number of \anables by a smgle
vanable under two drametrrcally opposuteconditions In the rst,
the variables haxe each the same effect (except 130551nyfor scale
drflerences nhlch dlsappear upon redefimtron) upon all of the eco
nomtc functions under discuss1on In tlns case they can bc
summed and treated as one Thus, we might dene as the same
commodltles all goods Inch h'ue exactly the same Influence on
consumption preferentx elds and production functlons \\ luch are
so to speak 1nmtelysubstxtutable ln practice ngtd adherence to
perfect substxtutabahty zmght lead to the unmanageable result that
; tno things are qwte the same Consequently, goods which do
TRANSFORMA TIONS, COMMODITIES, R '1TIONING I45
not signicantly (for the purpose at hand) differ are treated as
identical.
Actually the above case of substitutability is but one example
of the more general mathematical theorem whereby the identical
vanishing of a Jacobian, or of all of its minors of a given Order,
implies that there exists functional relationships between sets of
the variables, so that many may be dropped out of the picture
completely. It is clear from the implicit function theory that such
simplications are possible only in the case where the original
system was indeterminate. This may or may not be a matter of
concern. What if the variables which are extraneous cannot be
given determinate values by the system of equations which dene
equilibrium? The indeterminate variables may be a matter of
indierence to the economist. Thus, take any determinate system
of economic goods. Let each unit of one type of good be marked
with an invisible serial number, and now let us ask how many odd
numbered units will be bought by a given consumer. Clearly the
answer is indeterminate, but also of no possible interest. If con
sumers did have preferences as between different serial numbers,
then the indifference schedules would be affected by this fact, and
additional equations would be available to determine the nal allo
cation. And notice that even in the case of complete indifference
the nal allocation is not really indeterminate on any particular
occasion. It is just that the determining factors, which are taken
as due to chance by the economist, would be of a different
character.
Unlike the rst condition under which variables are lumped
together because they are innitely substitutable, at the other ex
treme we combine variables which bear an invariant, perfectly
joint" relationship to each other. A notable example is provided
by the classical dose of labor and capital applied to land. Here,
too, the economist customarily relaxes the rigid requirement of
perfect collinearity in favor of an approximate condition. _ThuS,
the most primitive justication for a price level constructols pro
vided by the indisputable fact that prices do generally rlse and
fall in about the same proportions.
" This must be qualied. Our original equilibrium equations may be imbedded
in a still wider set of equations so that the total is determinate, but the original subst
taken by itself is not.
I46 FOUNDATIONS OF ECONOMIC ANALYSIS
Geometncally thls dn'ers from the rst case m which 1nd1ffer
ence cun es or ISOquants are straight lmes Here the contours are
broken lunesmeetmg at rlght angles m such a way that the goods
or factors wnllby obvxous ch01cebe combmcd together 1n the gwen
proportlons almost regardless of price ratlos Unlike the chksran
bundle of goods the physxcal composmon Is the same regardless of
relatwe puces
Of all composxte magmtudes perhaps the most mterestmg to
the theonst xsthat of an index of the cost of hvmg, or of an index
of consumptnon output Such an mdecISdesxgned to meet certam
specxal requlrements and need not be satisfactory for other pur
poses In partxcular, lt need not represent the deszderatum from
the standpomt of the questions whlch Jevons and other proneers
m the use of Index numbers sought to answer Yet the theory of
these 1nd1cesISof some mterest for 1125
own sake and because ln the
course of mvestngatxon of these Index numbers econormsts mad
vertently stumbled upon ccrtam ordmal relationships whlch are
basm to welfare economics and to consxstent consumers behavmr
TY
MMOG
scoono
1|
FMS? COMMIY
FIGURE 1
locus and the axes For all such pomts were available to the m
d1v1dual m the mltlal Sltuatlon and he nevertheless preferred to
select X It follows that for all such pomts
U(X) : U(X) (67)
Thus we have narrowed down the region of our Ignorance but not
completely
It mll be noted that the two sntuatlons have not been treated
symmmxutally A always,berm; the reliera-m.mm. It might well
happen that X lay made the budget hne of the pomt X Assum
mg for the moment that knowledge of the complete preference eld
were avatlable for what values of X" would tlns be possnble? Ap
plying the requu'ement that X must cost less than X" at the prices
Pb the boundary of this region Wlllbc gwen by the equahty
z PX= z Mb (68)
TRANSFORMATIONS, COMMODITIES, RATIONING I49
Geometrically, this locus is the result of swinging the budget line
through the point X , and determining for each position the point
where it reaches the highest indifference curve, to which it is
clearly tangent. If we connect all such points, we have a familiar
oercurve. All points above this offer curve, are denitely better
than X , in the same sense that all points under the budget line
of X are denitely inferior to X. (It is to be emphasized that a
knowledge of two points does not give us this locus in the way that
it gives all the previous boundaries. Nevertheless, given two such
points, it would always be possible to decide just where X" was
with reSpect to X .)
We have now narrowed our ignorance down still further. Ac
tually this is as far as we can go on the basis of the given data.
Note that the old boundaries of our ignorance derived from the
utilization of quantity data alone have been made obsolete by the
additional light furnished by the auxiliary price information.
We have gone a long way, but there is still left a denite area
of darknessthe space between the two heavy lines. I should
like to state as strongly as possible that this nal indeterminacy is
intrinsic and inherent. No amount of ingenuity can remove it,
grounded as it is in the fundamental convexity properties of the
indifference eld, or more accurately in the consistency behavior of
the individual. It is important to prove this rigorously, for pe
culiarlrin the literature of index numbers is an attempt made to
search for limits within which the truth must lie without at the
same time investigating whether or not these are the best possible
limits. Moreover, the limits themselves are sometimes derived
under special approximations, such as the neglect of squares of
small quantities, etc.
To see that these are indeed the best possible limits under the
circumstances, let us suppose that someone proposes narrower
limits. Since our preference eld is arbitrary except possibly for
certain curvature properties, we can draw the true indifference
curve through X so as to contradict any more denite result. if
the proposer claims that a given point in the area of darkness IS
worse than X , we can pass the indifference curve below that point
but above the correct lower boundary in such a way as to make this
statement wrong. Similarly, the opposite statement can be shown
not to hold universally. (It is of course understood that we do
150 FOUNDATIONS OF ECONOMIC ANAL you
not actually change the Indifferenceeld of the economlc umt under
observatxon, but we" can nd a consxstent eld for whtch the
gwen result holds ) Where the offer curves bend back we are able
from quantity consrderatnonsalone to extend our boundanes along
the dotted lunes lndxcated 1n the gure
For well behaved Indifference elds wrth the proper contmuous
denvatlves, the offer curve whlch gwes the upper boundary of the
area of darkness Willbe tangentral to the mdnerence curve gomg
through X and to the lower budget boundary through X Even
If there xsa corner the Indifference curve at X ". there must be
generalnzed tangency the sense of touchmg from above wnthout
crossmg One mlght be tempted to say that the area of mde
termmacy narrows down to a pomt m the vxcmlty of X, but
nothmg IS gamed thereby Nevertheless, the mathemattcran :s
tempted to conSIder Taylors expansnons ln the neighborhood of
the X pomt and to neglect terms of hlgher order Perhaps there
IS even some emplncal stat1st1cal usefulness 1n the handllng of
budgetary data to be found m th1s practlce, Wthh has been assoc1
ated With the names of Bowley and Wald
But from a fundamental vrewpomt there lS no gettmg around
the fact that for any mte move, howeversmall, there remams a
regzon of 1gnorance, th1s reglon may narrow down as the srze of
the movement decreases, but xt never vanishes, except for the
tnvral se of a vamshmg movement Thus. there ISnothing t
be gamed for the present purpose from followmg the practtoe of
DlVISla" m workmg thh dnerentlals or rnmtesrmals The fact
that these avo1d dlcultles connected wrth the tlme reversal and
factoral reversal ISnot an 1nd1catlonof thexr superiority, but of the
fact that they srdestep the mtnnsxc dxcultles of the subject
matter
By the exammatxon of value sums we can somet1mes, but not
always, state denitely whether one sxtuatlon ls worse than another
Butwe can never by these means State 't'na't two sttuatlons are
equally desnrable In fact, as wnll become evxdent from later dlS
It may be worth mentioning that even a knowledge of the curvature of the m
da'erencecurve at X wallnot enable us to narrow down our reglon ol Ignorance It
mll enable us to assert wrth condence that the mdn'erencecurve wall approach arbt
tranly near to the escalating cucle, but the departures for any mte small movement
may be of either 5130and of any magmtude
"' F Dwzsra, EconormqueRatwmllc (Pans 1928)
TRANSFORM'A TIONS, COMJIIODITIES, RA TIONING I 51
cussion, knowledge of prices and quantities on a nite number of
points will not permit us to determine the equality of two points;
but in the limit as the number of points becomes innite we can in
fortunate circumstances determine points of indifference.
This result may be approached by a consideration of the addi
tional information which a third point will yield. We now have
three pairs of points, and it may turn out that any two of them
when put to the test of our computed value sums gives a denite
answer. In that case each pair can be considered by itself without
regard to the third point. Of course, from the discussion of chapter
v it will be clear that a consistent ordinal preference eld can never
give contradictory testimony whereby point A is better than B,
which is better than C, and at the same time the latter is better
than A. For the relations of ordinal utility are transitive.
But the relations of better or worse as revealed by the value
sums are not transitive; that is why I have elsewhereproposed a new
notation to represent revealed" preference in this special sense.
Thus, if our value sums give a denite result so that
2 PX }: PX, (69)
this fact may be represented by the symbol
X " @ X . (70)
Since
X" @ X implies U(X") < U(X), (71)
and this in turn implies
U(X) <1:U(X), (72)
then to avoid a contradiction in logic, we must be able to state
the theorem
X"@X implies X@X".
But this is not the same thing as the meaninglessassertion that
Xb @ Xl implies X X", (73)
or the incorrect assertion that
X @ Xl and X @ X" implies X @ X. (74)
The most that can be stated under the above hypothesis is that
X @ X. (75)
152 FOUNDATIONS OF ECONOMIC ANALlSIS
Thls IS much weaker than transrttvrty An additional Indicatxon
that the algebra of revealed preference ts quite drstmct from that
of cardinal or ordmal numbers 15 the fact that equahty 15not
dened consequently, two pomts cannot be placed m one of the
three categories A worse than B, or B nor-se than A. or the tno
equally good All we can say IS either A :s revealed to be worse
than B or B 15rexcaled to be norse than A or there ISno Indlca
tron one way or another These are mutually exclusne categones
only If a consrstent preference eld ls postulated
It rs precxselybecause of this lack of transnthty that knowledge
of a third pomt may add knowledge to a companson bet een two
gwen potnts Our value sums may gne no Inchcatton wzth respect
to pomts X and X taken by themselwes but an Intermedrate pomt
X may serve to Indicate their true ordinal relattonshlp to each
other We are m a posmon now to mchcate Just how much pomts
X and X may narrow down our Ignorance relationshlp to X
If the two addlttonal pomts both he m our old area of darkness
ne are no better off than before But nfX hes m the upper regron
of certamty w1th respect to X and X hes 111the upper region of
certamty wrth respect to X. then even though X hes m the X' s
reglon of uncertainty It can stnll be demtely sand to be better
than X Therefore we have narrowed down our area of darkness
How much can thrs area be reduced the most favorable case?
From the geometry of the problem It can be shown that )1"5must
he on X s o'er curve for the best results We then proceed to
draw the offer curve through X Tlns wnllcross the old one and
thereby narrow our reg1on of Ignorance since every poxnt aboxc
the new offer curve 15denitely better than X If now ne let A
take each and exery posxtronon the old oer curve the process can
be duphcated as many tlmes as we msh, gn mg us a one parameter
famlly of new offer curves The lower envelope to thns family of
curves gwes us our new upper boundary of uncertamty
In the same way, ne can secure a best lower boundary by lettmg
X" travel along the ongmal budget lme, generating through each
such pomt a new budget lme or a one parameter famxly of such
lunesm all The upper envelope of ats farmly of lmes ts our new
lower boundary Our new boundanes must of necessrty he wrthm
the old ones but there ISstlll left an area of darkness
If a fourth pornt lS added we may narrow our Ignorance stnll
TRANSFORM'ATIONS, COAIMODITIES. RATIONING I 53
further; similarly with a fth point. It is intuitively clear that
the boundaries will never meet for any nite number 0f POntS:bUt
in the limit as the number of points becomes innite, the upper and
lower boundaries approach a common limit, which is of course the
indifference curve through X . Points along this curve, and only
such points, can never nd their place in a nite chain of points
which serve to relate them unambiguously to the initial situation.
Thus, with a single dimensional innity of points any indifference
curve may be traced out. With a two-dimensional innity of
points, the whole indifference eld can be determined. In fact,
with this much knowledge we may dispense completely with re
vealed preference, and instead integrate completely our elements
of slope at each point into a one parameter family of indifference
curves.
The above discussion is concerned with the narrowing of our
eld of ignorance under the best circumstances. in actual practice
we shall not necessarily do nearly so well. An instance is provided
by the frequently met case in which we know a complete budget
line (expenditure path, income consumption curve, etc.) giving the
behavior of changes in all of the commodities with changes in
money income but with unchanged prices. Such a situation is
provided by observations of the behavior of more or less similar
people, all confronted with the same prices but differing with
respect to total expenditure.
Not only is this of considerable statistical importance, but it is
of special relevance in connection with the usual economic theory
of index numbers, where a comparison is made between two price
situations, rather than between two particular price-commodity
situations. By applying our previous analysis it will be seen that
the knowledge of an innite number of points along two budget
lines will not be sufcient to determine the whole preference eld,
or even a single indifference curve, or even be sufcient to enable
us to match points of equivalent satisfaction on the two budget
lines. If we select a given point on one line and through it draw
a budget line and an offer curve, then these last two loci will split
the other budget line into three parts. The lower part will consist
of points all worse than the initial point, the upper part will consist
of points all better than that point, while the intermediate part will
constitute our area of indeterminacy.
154 lUUlVUJJUdUF thztvtulu nzvnuluzo
Of course the closer together are the two prxce situatxons the
smaller w11!be the mdetermmacy Therefore we can narrow down
our Ignorance by a knowledge of many Intermediate budget lines
In the hunt as the number of budget lmes becomes Innite such
a way as to decrease 1nden1tely the distance between each we
shall approach the complete Indifference eld
The value sums conSIdered up lll'ltll now have all had the d1s
advantage of bemg capable of yleldmg mdetermmate results We
shall put up wath this If need be but rst we must dctermme
whether xtISnot possrble to devnsean Index of quanttty computable
only from the prxces and amounts of all the commodities which
W111 be an unfathng 1nd1catorof ordlml utthty Can we not nd a
magic formula wh1ch wnll have these properties when apphed to
any preference eld whatsoever or at least to those of the proper
convemty>
The answer 15no As yet no one has dewsed such a formula
and the followmg mathemat1cal reasomng shows why no one exor
can even xr we permit more general functlons than snmple value
sums Any such general formula wrll from the nature of the
problem be a functxon of quantltres and pnces m one or more
(usually two) s1tuatxons Without loss of generahty we may con
srder the puces and quantltzes of all but one of the Situations as
xed whlle the one sxtuatlon represents marbltrary varnble poxnt
In terms of ltS pnces and quantltres (P A) we must be able to
construct a smgle Valued function Q which IS constant so long
as the X pomts remaln on the same mdl'ercncc cune of the
(unknown) preference eld Because this xsto be a quantxty 1ndex
xtxsclear that relatu e prices only can be Important and so we may
express all prlces m terms of the rst good as numerau'e Then Q
may be wrxtten as
= 6 ,, - 76
Q (at; x P:
P1 pa
P1 ) ( )
For observable equnhbrlum quantntles the ratlos of pnces are equal
to marginal rates of substltutnon (U?2 1R") or to ratros of
margmal utxhtles (U2/ U; U../ U;) Although the form of the
relationship wnll change from preference eld to preference eld
and IS m any case unknown to us there W1llstlll nevertheless ln
h specxc mstance be a functlonal relatxonshlp between these
TRANSFORMATIONS, COMMODITIES, RATIONING 155
ratios and quantities of goods. Therefore Q will be a uniquely
determined function of the x5 alone. In fact, since Q 13constant
along an indifference curve, it must itself be one cardinal index of
utility; its partial derivatives with respect to each commodity
(computed by adding to the direct effect of each x on Q all indirect
effectsvia inuence on price changes) can be interpreted as marginal
utilities, whose ratios are equal to observable price ratios, etc.
In consequence of this last fact, equation (76) can be written
in the form
a_Q_ _
9(x11'. ,x", axl/x21"'yxl __)_
xn Q= 0. (77)
This identity is a rst order partial differential equation which the
preference eld must satisfy, and only a limited small subset of all
preference elds will do so. Consequently we have proved the
impossibility of nding a magic formula to serve as a quantity indi
cator in the general case of a consistent ordinal preference eld.
In actual fact it is customary to restrict the form of a quantity
index still further, to require that it be a homogeneous function of
the rst order in the component quantities, etc. This restricts the
range of applicability still further, to preference elds possessing
expenditure proportionality, etc.
For any particular 0 it may be a problem of some mathematical
difculty to determine the exact restriction on the preference eld.
But where 0 represents the Laspeyre quantity index, thought of as
a function of the quantity components of the second point, with
base point xed, it will be immediately seen that the only possible
preference eld for which it is exact is that of a family of parallel
straight line indifference curves. Of course, this is almost a re
ducto ad absurdum, since convexity is denied, and since all relative
prices (of goods which are bought) cannot vary.
The ideal index number is one of the most popular index
numbers. It is the geometric mean between Laspeyres and
Paasches index numbers. A. Konus and S. Buscheguennce,23as
well as S. Alexander (in an unpublished Harvard paper), have
shown that it is exact only for certain hyperbolic indifference
curves. This theorem may be derived by solving explicitly the
" See the reference in H. Schultz, A Misunderstanding in Index-Number Theory,"
Econamcm'ca, VII (1939), 8.
156 FOUNDATIONS OF ECONOMIC ANALLSIS
unphed dn'erentnalequatxon Similar results can be derrved for
other approxxmate" formulae Inch have been suggested
As a result of the analysrs of this section the conclusmn may
be ventured that the Important economic content of Index number
theory resrdes m the fact that xt attempts to utlhzc hunted pracc
and quantlty data to mfcr ordmal preference compansons The
aboxe formulatnon seems best desxgnedto reveal thts essentnl con
tent, and to show the mtnnsxc hmltatxons necessarlly lmolved
In the nest sectxon the same analysts rs shown to bc useful m
connection thh the more commonly met formulation of the Index
number problem
PRESENT Fommmnows or INDE'C NUMBERS
We are now m a posmon to apply our tools of analysns to the
more famlhar branch of Index numbers First, e consrder pnce
of hvmg comparxsons bet een tuo dx'erentpnce srtuatxons The
price of Iwmg ldCCnumber ln gomg from the srtuatlon (X') to
(X) rs famzharly dened as the ratno of the cost of the cheapest
bundle of goods at the pnccs of the second sntuatxon nhnch mll
yneldsatrslactron equnalent to that of the xmtzalSituation to the
cost of the zmtral bundle at the mltxal pnces
We assume as known (say from empmcal studies on Identical
xndrwdualsthh \ arymg mcomes) the pace expansron paths untuo
pnce sxtuatxons (P) and (Pb) These are dened by the followmg
sets of parametnc equations
a. = 3.(P1'. , p., I), (t 1, u) (78)
x : "*(Plbo vpast I) ( I: a n) (79)
Consrder an :mtxal srtuatron (X) on the Imtral expansron path
The 1nd1fferencelocus correspondmg to thrs pomt mil xntersect
the second expansron path m a p01nt called (X) such that
mx) = U(X) (80)
An ldCinumber of the prxce of In mg ISdened as follows
_ ): FX
105_ ___
PX (81)
Slmrlarly,
a be
pa = ___Z
PX
Z PX '
(82)
TRANSFORMATIONS, COMMODITIES, RATIONING 157
where
U(X") = U"), (83)
and (X) lies on the initial expansion path.
These respective index numbers are not rigidly related; in
general they are not reciprocals. Since we do not know the hypo
thetical preference eld, (Xb) is not known, and so our index
number cannot be computed.
However, from the considerations indicated above we know
that all of the points on the expansion path (b) fall into three
classes with respect to (X):
1. (X) @ (X) z PaX g z P0X (34)
2. (X) @ (X) z PbXa g z PbX (35)
3. (X) @ (sz) z PX > z PaXa (86)
(X) @ (X) ): PbXa > z PbX.
In particular, consider the equality signs in (84) and (85) and we
get the boundary points dened by the intersection of
2 PK = {: PbX (87)
and
x,- = hi(p1b, ' ', Pub, I). (Z = 1, ' ' ', n) (88)
F = 2:312;be (100)
subject to
%; s %, (102)
(*X) is easily computed since
*xmab= E PX (103)
pmu
Qa=ELJLZP
21wa
):?wi n
=F1? (2)
i.e., this index is the ratio between the actual cost of the batch
()F)to the cost of the cheapest bundle at (Pb) prices which would
yield an equivalent satisfaction to that of (X).
Similarly,
a }: PX I
Q"= W =FF (113)
Obviously, if one situation is preferred to the other, the index
number between them will be less than one, i.e..
U(X) < U(X") implies Q < 1, (114)
and conversely. Likewise
U(X) < U(X") implies Q > 1. (115)
Hence,
Q > 1 implies Q < 1, (116)
and
Q > 1 implies Q < 1. (117)
More generally,
Q = 1 implies Q = 1, (118)
where the inequalities are to be taken in the indicated order.
However,
1
Qab 75 ?; ! (119)
except under special circumstances. The above relations are
consequences of the concavity of the indi'erence loci.
Of course in the absence of knowledge of the preference eld,
it is impossible to compute these index numbers, since we do not
know (Xb)and (X). From the last section we know that
2 P xx :. >: Pb lx < }: PX< 2 Pbax. (120)
Hence,
PbXb PX" Z PX
%FFFF>Q>TFFF' (
162 FOUNDATIONS OF ECONOMIC ANALYSIS
or PUP ): P aXs E PX
?! = :: ! xb
> Q > zmr- (122)
As before. ('X) can be computed only If c know the cxpansmn
paths From (103) we nd
W
ZPX '
KW
p.: EPX (123)
Therefore.
p. }:PX'
Pa. E Pch "=- zXFX
Pblx > O> _);PX
2 P6X. n Paasche (124)
Smularly,
? [9,345,
g I: ? Px'xi E I" c }; Pfx: 5--I' (128)
% (p,-+ )Ebkp,")x,-
= I + gap. (134)
Actually, from a welfare point of view it can be shown that the
free interchange of different kinds of stamps, among themselves
and against money, is in a certain sense optimal. In fact, it is an
indirect way of permitting individuals to exchange goods so long as
the exchange is mutually advantageous. These last two sentences
must be qualied. They must not be taken to mean that every
individual will be better off if coupons or goods can be interchanged.
Thus, if the rich buy redundant ration coupons from the poor to
the mutual advantage of both, their expenditure by the rich will
cause an increase in the point price of the scarce rationed goods in
question to the disadvantage of the middle classes.
Nevertheless, it can be shown that free interchange is optimal
in the sense that its introduction accompanied by appropriate
modications in the point allocations to every individual could
lead to an improvement for everybody. In the above examp'e,
the middle classes could be bribed into acquiescence, and there
would still be a margin of advantage left for rich and poor. I do
not imply by this that the middle classes shouldreceive such a bribe
since that would suggest a belief in the perfection of the previous
status qua. Nor should it be thought that anything said here is an
argument for making coupons interchangeable, since there might
very well be in actual fact very grave difculties in the way of
devising a method of point allocation which would recognize the
harm done to particular individuals.
CHAPTER VII
_ . F = F(so), (3)
for which
62F _ 0
xy : ' (4)
Successively differentiating (3) partially with respect to x and y,
we get
62F I II
a??? = F 99111
+ F (Prior (5)
This, together with (2), requires that
F "90:99E 0, (6)
or
F"(sa) E 0- (7)
Therefore,
F = a + I790, (8)
where a and b are origin and scale constants respectively.
It is clear, therefore, that the assumption that utilities shall be
independent will help to select one utility index as the cardinal
measure of utility. Nevertheless, even this convention is not in
general applicable. It will guarantee us that we do not have two
different utility scales, as has been shown in the above proof; it
will not, in general, provide us with even one scale.
If We assume an indierence eld obeying the ordinary con
cavity restrictions and nothing more, then there will not, in general,
be even one utility index which can be written in the form
90: E 0
Rd; + d =%x;
dx + d) (18)
[ hrs can easuly be transformed mto the ex tct dlffercntnal
dp = h(x)dx + k(y)dy. (19)
SPECIAL ASPECTS OF CONSUMER THEORY I77
OI
x_m
_ PvP"
(.13.: JL)
(24)
y = n (_P 12:
These demand equations must be subject to the restnctxon
2log %
_y:
xy '0 (25)
When there are more than two goods the restncttons nmphed
by the very poss:b111tyof an Independent Index of utlhty take on
a dlerent and more complicated form If there ex15tsan underF
for Wthh
Fu Oo (" # .?) (26)
then
F : F,(P)P) + F(P)Pt: E 0: (27)
where p15any other Index Thus
(at! =
___... 28
M Tu) (a # J) ( )
where T xs an arbxtrary function, and go15 any mdex of utlhty
Takmg mto account the (n 1)(n 2)/2 condltzons of mtegra
blhty th1s xmphes an addlttonal 11(n ~1)/2 condlttons It ISto
be noted that these are ldentmes holdmg everywhere Not only
are they necessary, but the transformahon
F = f e'TWdv (29:
shows them to be sufcnt as well
In terms at the mdtieteme var.-wma.these. take. std]. anothel
form Let
2-,.= .. & __;1
Pr _ dx: __ R(x1 , x,.) (30
These, of course, satxsfy the Identttles
{Ri
r 5 *R (2.1.12 = I. .n) (31
SPECIAL ASPECTS OF CONSUMER THEORY 179
If an independent index of utility is possible, then we must have
a "R5
xk 20. (jki) (32)
In view of the n3 relations of (31), the above n(n 1)2 relations
are not all independent. At most 1z(n2) are independent, and
these may be written in the form
alRi . o I .
xJ. 50' ($95.7)! (?,_7=2,---,n)
lRi (33)
a(12) .:..0_
___1_
6.171
(i = 3, ...,)
These conditions are both necessary and sufcient. They im
ply among other things the (n 1)(n 2)/2 integrability condi
tions of equation (14) of chapter v. On the other hand, if the
latter are postulated at the beginning, then equations (33) cease
to be all independent and can be reduced in number.
We have then n(n -2) partial dierential equations of the rst
order. Subject to an equal number of arbitrary functions, the
general solution is uniquely determined. But empirically an ob
servation of an expenditure path involves (n 1) functions.
Hence, observation of more than n(n 2)/ (n 1), or more than
n expenditure paths, could be used to disprove the possibility of
independence.
In words, the implications of independence are that the amount
bought of any good x,- can be expressed in terms of its price p,,
the price of any other good {J,-,and the amount of expenditure upon
these two goods alone, i.e., upon the amount left over out of total
expenditure after all other goods are bought. Thus, the general
demand function
xi = ]?an ' ' , imI)
can be written in the special forms
(Pn ' ', Pm 1) E hijbi, P131 + Pax.-+ [),-x,- >; kak) (34)
for any i not equal to j. Subject to the conditions of (25) and
certain consistency relationships, these conditions seem to be
sufcient as well as necessary.
4
I80 FOUNDATIONS OF ECONOMIC ANALYSIS
Except for R. G. D. Allens derivation of equation (15) for the
two-good case. 1 am not aware that these full price-quantity impli
cations of independence have previously been derived. However.
fragmentary sets of necessary but by no means sufcient conditions
for independent goods have been derived by Slutsky and by Milton
Friedman They differ fundamentally from those above, and are
of a species which I choose to call local implications of independ
ence. At a given pomt in the (X) space it may be possible to nd
a transformation F(<p)such that
Fu : FIPU+ F"q0,qa,E 0 (" # ]) (35)
Thus, if there are only two goods (xx) and (x2), this is always
possible at every point so that the local conditions degenerate into
trivial identities At any point (X) = (xi, are),let
F"() _ Mei-Y)
m) mxwxr
and
Flo,(X) = 0. (37)
Pu : (" A (22
P " 0? pmPn+1") ! (39)
R. G D Allen, A Comparison between Different Denitions of Complementaf
and Competitive Goods," Economstnca, II (1934} 168-175
' Milton Friedman, "Professor Plgou's Method for Measuring Elastique: of Demand
from Budgetary Data," Quarterly Journal of Economws. November, 1935. pp 151163
E. Slutsky, Sulla teorla del bilancm del consumatore," Gromale degli Economist!
LI 915), 23-26
SPECIAL ASPECTS OF CONSUMER THEORY 181
where
6x1
K11 K12 ' KI" 6]?
636:
K12 K22 ' ' ' K2n
p = . . . . ' (40)
8 n
Kln K2n ' ' Kan 'ai]:
a] a] 61
and the subscripts indicate appropriate cofactors. Also
a 1
09 = .(I) (41)
in the notation of chapter v.
If an index exists such that
F,-,-E 0, (i # j) (42)
at a point, then
E
"h'i = (Ml _ km,r
m': 1 km,[)+ 2.1[(k5
1m '- kt)1 "'_ kikj(77i'77j)]_
km,r (44)
In the two-dimensional case the restriction is vacuous.
1Friedman, op. cit., p. 162.
182 FOUNDATIONS OF ECONOMIC ANALYSIS
From the fact that their number is excessive, it is clear that
they cannot all be independent. In fact, I am unable to prove
that they are complete,i e , equivalent to the above Slutslty condi
tions; indeed, I would venture the conjecture that they are not
It would be literally imposable [or any indiv1dual to determine
by introspection whether or not his demand functions satised the
above relations Indeed, the likelihood that such relations, se
lected arbitrarily from an innity of possible functional relations,
should hold is innitely small The minute amount of plauSibility
" On page 162, footnote 2, Mr Friedman argues against the admissibility of the
Pigou theorem that
k k
Er when =0= (o)
1 "!!! mt '!1!
on the grounds that " there is a presumption that 7mand k; [my notation] are
"it truly related " The whole problem is of course ambiguous until a particular set of
indifference curves are speCied Nevertheless within the realm of probability and
presumption, Mr Fnedmans statement seems to be incorrect In chapter v it i:
shown that
I
Z 12.170
1
=1.
Z k.
I
i e . a weighted average of all income elastiCities equals unity regardless of the number
of goods The average k is given by
Z le.
, = ..l
n n
This approaches zero as the number of goods increases
A survey of empiriml budget studies Will convmoe the reader that the incomi
elastiuties are distributed around unity, while the proportions vary around 1/1
See also A C Pigou and N Georgescu Roegen, Marginal Utility of Money ant
Elastique: of Demand. Quarterly Journal of Economics,vol L. no 3 (1936)
The above Pigou theorem rests, nevertheless upon the restrictive assumption 01
independence If it could be shown that it holds for small values of le./rm in ever? 0355
it would be much more useful
That this cannot be true in general is shown by the followmg speCial case Lei
consUmption of n goods be completely JOInt in the sense that there are always xez
prOportions Then regardless of the number of goods m: = 1. and via/1m = 1 0
the other hand it is easin shown that ii i E np. Consequently. flu/n = ku/h
Even though we make the k : become arbitrarily small as compared to the income 9135
nana, their ratio may take on any value other than unity Thus, the Pigou theoren
() is false
Oi course. this example rests upon discontinuities in the higher derivatives. but l
should thin]. that e could approximate this condition by appropriate chance of con
tinuous derivatives
ornblb AOIDDIO UF LUIVOUJWK
J'HUKY 153
to the proposition that independence is an admissible assumption
comes from the fact that the question is usually posed in such a
way that independence appears to be an intermediate class between
the extreme classes of competitiveness and complementarity. Em
ploying a rudimentary concept of the equal probability of the
unknown, one is inclined at rst blush to acquiesce to the inde
pendence assumption. The error involved in so doing is obvious
from our previous discussion, completely aside from the crucial
ambiguity attaching to the older notions of independence and
complementarity.
COMPLEMENTARITY
In my opinion the problem of complementarity has received
more attention than is merited by its intrinsic importance. Never
theless, as a result of the interest in this subject, crucial incon
sistencies in the thought of Pareto were revealed by Hicks and
Allen, and much light was thrown on the cardinal and ordinal
conceptions of utility.
The older writers, Fisher, Pareto, and Edgeworth, suggested
as a qualitative denition of complementarity between two goods,
9:,-and x,-, the sign of the cross derivative of the utility function;
i.e., goods were complementary, independent, or competitive de
pending upon whether
62 > = . 45
xixj < O ( )
If one assumes only an ordinal preference eld, all numerical utility
indexes are equally admissible. This cross derivative is not in
variant in sign under a change in the index of utility. Consider
a monotonic transformation of a.
dF
U ), d, > O
U;: _piy (46)
dF sz
U;,- = (170 a.-,-+ dPz sa.-soi
It may be argued that regarded purely as a working hypothesis the facts do not
sharply contradict the independence aSSUmption. A little investigation revls that
such a hypothesis has not been tested from this point of view. On the contrary, it is
implicitly assumed from the beginning in the manipulation of the statistical data.
Hence, one would have to go back to examine the original empirical data. It 15interest
ing to note that observations on three expenditure paths would be sufcient to contra
the independence assumption in the case of two commodities.
I84 FOUNDATIONS OF ECONOMIC ANALISIS
By a proper selection of the arbitrary term, sz/dlpz, the Slgnof
the new cross derivative can be made to differ from that of the old
As a corollary of this lack of invariance, complementarity so dened
Willhave nothing to do with the budgetary habits of individuals
With respect to the two goods in question. Similarly, an indi
iiduals introspective, spontaneous, intuitive designation of tuo
goods as complementary or competitive (sugar and coffee.beefor
pork, etc ) corresponds not to such a measure as this, but rather
to behavioristic properties of the preference eld and demand
functions
In the course oi the last xe years the world's best economists
have spent conSIderable time and energy in the study of Professor
Hickss Value and Capital It is perhaps symptomatic of the
essential unimportance of the concept of complementarity that in
this period no one seems to have noticed that the author givesmo
or more distinct (and inconsistent?) denitions of complementarity.
The verbal definition in the tent (p 44) seems to be di'erent from
that in the mathematical appendix (paragraph 8)
Hicks and Allen. SlutsLy. and Schultz have suggested invariant. measures ol
complementarity which are properties of the indifference curve system and demand
function Perhaps the Simplest measure of complementarity between two goods 24
and x,. is the sign of
31'. az. K"?!
K = - : = K =
ii ap, + , aI .. or m x;
If there are but two commodities, this must always be posutive in Sign. In the many
commodity case at least one must be of positive Signin order to satisfy the relationships
2. 93K = 0, (. = la | n)
K.. <O
Professor Leontiel has suggested the followmg invariant measure of independence
corresponding most closely to the old notions
.
azdR: , lg( U, )
ax; ' x.:|:, _ 0
Since the above lines ere rst written Professor Hickss two works. Valueam
Capital and Theam: Mathemahqzde La Valeur, have appeared I still believe. however
that Hickss own solid contributiOns to economic theory do not rest on his treatment oi
complementarity, and that the extended discussion of the concept is a tribute to an 01!
love rather than the necessary consequence of the subject matter
J R. Hicks, Value and Capital (Oxford, 1939)
Actually. there is still a third denition in the footnote on page 44 Only I "
case of three goods is it clearly equivalent to one of the other trio. in the general 035"
SPECIAL ASPECTS OF CONSUJIIER THEORY 185
This may be seen in many ways. While the mathematical
denition can be applied to the case where there are only two goods,
the literary denition cannot. (It is irrelevant that in the case of
two goods the mathematical denition of complementarity admits
of only one possible algebraic sense.) There is no reason why two
distinct denitions should give the same answer in any particular
case, so it is not surprising that one can invent examples ad im'ni
tum for which two goods, such as wheat and linen, are on one of
the denitions complements and on the other substitutes. But as
we shall see in a moment, things are in even a worse state. Ac
cording to the denition of the literary text, it is possible for wheat
and linen to be complements and substitutes at the same time,
depending on the selection of the third good which is to serve as
numeraire. The denition is ambiguous; instead of reecting the
properties of two goods, it (or rather they) represents the properties
of three goods.
The mathematical appendix denition gives as a coefcient of
complementarity between the ith and jth good the element of the
ith row and jth column of the matrix ($,-")where
x = #19, U) (47)
represents in matrix notation the demand functions, for each
indifference level.
Because of the homogeneity property discussed in chapter v it
is clear that these n demand functions cannot be inverted to give
each [2 in terms of the x's and U. Only relative prices can be
determined. We can use any one price as numeraire, discard one
of the above equations and invert the others so as to give (n -1)
relative prices in terms of (n 1) goods and the ordinal level of U.
It is most convenient to omit the quantity of the numeraire, .17.,
and its demand equation. This gives us
%
1: = Wie-+. -- -. x. se. ---. x... U). (+ # n) (48)
is ambiguously worded, since the o'er of the maximum amount of money by the con
sumer to attain given amounts of the two goods in question will change the amount
bought of all other goods. It wouldappear that the author intends that all other (than
the two goods in question and the numeraire good) goods are to be held constant, in
which case it becomes identiml with the verbal denition. The difculty may reside in
Hicks's unique use of the concept money."
186 FOUNDATIONS OF ECONOMIC ANALYSIS
Let (*Nj.) = N; be the matrm of thrs transformation Then the
element corresponding to the 1th prlce and 3th quantity represents
the literary demtron of complementanty It can casxlybe shown
that thlS, hke the rst matrix, must be symmetncal In fact. xt
15necessanly negatwe demte, hke the gbmatrn.
These last two statements are consequences of the demtlon of
the Inverse demand funct1ons accordmg to which
Nl: : kIUJln): (49)
AxABj a o. (51)
If the interest rates for all periods are equal, and if we go to the
llmlt, we have the denite theorem
wxj/dmjafl % 0. (52)
_ f p a xidpio
3. The amount by which the expenditure on the 13thgood in
the new situation is exceeded by the maximum amount of money
which the consumer would be willing to pay for x.-"in preference
to trading at the old set of prices. (This may be negative if we
are dealing with a price increase rather than a decrease.) Call
this bEab.
4. The amount of extra income which the consumer would in
sist upon if he is to be as well o' as in the new situation while
consuming the old amount of xi. Call this Eab.
5. The change in income which will make trading at the new
set of prices as attractive as trading at the old set of prices with
the initial income. Call this bAIab.
. The change in income which will make trading at the old set
of prices as attractive as trading at the new set of prices with the
initial income. Call this Aha.
According to the Marshallian doctrine of consumer's surplus,
all six of these magnitudes are equal except for dimensional con
Interchanging subscripts changes algebraic signs. Thus,
Alas = MI. AI, but not "AIabnor 5E, can exceed I.
200 I'OUNDATIONS OF ECONOMIC ANALYSIS
stants We are expllcxtly warned, however. that hxsdoctrme holds
unqualxedly only when the marginal utnllty of mcome ISconstant
and only If utllmes are independent I shall now examme the
value of each of these magnxtudes four cases (a) m the general
unrestncted case of Stable demand, (b) under the rst mterpreta
non of constancy of the margmal utxlxty of Income, (c) under the
second hypothcsns when the 1th good 15 not the numerazre, and
(d) under the second hypothesns when the zth good Itself has con
stant margnnal utlhty of meome Only the most sketchy proofs
w1llbe Indicated
In the general case we have the followxngrelations
- -f,()f()
=jptm.$(p,-f)dpl= _ffmgdp. (56)
AI... = max (2"; pf'xf Z:p,"x,), where qa(X) = rp(X). (57)
1 z
EW - Mac; (591
a ab
mm 0:: m, -
)::prxr) where par) = saw) (611
W1 5U 1
+ U2 + ---, (S)
where as usual subscripts represent partial differentiation, but
where the superscript is taken to indicate different individuals, we
shall see that this has the dimensionality of the good, 0, and
nothing else.
Pareto attempts to show that if the original position is one of
equilibrium under perfect competition, then no possible variations,
consistent with the fundamental scarcity of goods and given tech
nology, can make the above expression positive. If it could, he
says, it would be possible to arrange things so that each term in
the expression could be made positive, and then everyone would
be better off. But the expression cannot be made positive. Ac
tually, regarded as a differential expression (of the rst order), the
above expression can be shown to be zero, in consequence of the
fact that each ,commodity is sold at minimum unit cost (propor
tionality of marginal products, etc.,11and in consequence of the
tangency of each individuals indifference curves to mutual price
exchange loci. If recourse is had to differentials of higher order,
the secondary maximum-minimum conditions of rms and indi
viduals will guarantee that the expression (5) will be negative for
all nite deviations from the competitive position, or so Pareto
attempted to show. _
Although Paretos treatment is somewhat sketchy and in need
of expansion, part of which Barone later providedn the mam out
lines are reasonably clear. But in connection With the funda
u Ibid., p. 646.
214 FOUNDATIONS OF ECONOMIC ANALYSIS
mentals of Interpretation of the Slgmcance of his maxxmum, there
do anse certam problems Fu'st, can the dlfferentlal e\pre5510nof
(S) be regarded as the exact drerentlal of some etpressmn? Ac
tually, Pareto later glves a name to th1s expressmn, calling 1t U,
but IS there an expressmn U (soc1al utlhty?) of whtch thrs 15an
exact differentlal? Pareto does not tell us, but presumably he
would answer. no, if he were on guard when the questlon was asked
As we shall see, Barone does work 1th an expresswn whose dl'er
entxal corresponds to (5), but he clearly recogmzes that 1t xs a
constmctron, not Im.olvmg the dtmenswn of utlhty, but rather
that of the numeralre good
But the most Important objection to Paretos exposntlon 13his
lack of emphasxsupon the fact that an optlmum pomt, lus sense
:s not a umque pomt If transfers of Income from one 1nd1v1dual
to another are arb1trar11y 1mposed, there Will be a new optimum
pomt, and there 13absolutely no way of decxdmg whether the new
pomt lSbetter or worse than the old H15optimum pomts const:
tute a mamfold mmty of values Thus locus can be obtamed
under reglmes qmte dlfferent from perfect competxtxon (e g, by
multilateral monopoly) \Vnthln Pareto's system 1t 151mpossnble
to decxde, by hlS dl'erentlal cntenon or otherwnse. which of two
pomts on what may be called the generaltzed contract locus" are
better, or even that a gwen movement off the contract locus and
hence to a non optimal poxnt 15good or bad Actually m terms
of the under reference schemes of ordmary economlc thought such
a movement may be deemed emmently desnrable But Pareto
shows that however deszrable such a move may be, there emsts st1ll
a better move, which for the same (ordmal) amount of harm to
those Who should" be harmed. Wlll yzeld more benet for the
worthy ones who are to be beneted Thls xs an 1mportant
contrlbutlon
7 1.9a masterly artmke, mrrttas- m 1503 .? kalam bat m!
translated mto Enghsh unt 1935,laBarone developed further and
m greater detanl the Paretlan condltrons of optimum. especnallyas
In hls rher Cour: dtscussxon II 90 H he exphcntly assumes the dustnbutmn of
Income to be convenable but m 1118later Manuel treatment the dependence of the
optimum pomt on the mmal dustnbutxon of Income and hence Ita lack of umqueness
:s not brought to the foreground
" Reprmted as an Appendlx F A Hayek ed Calledwut Economic Ptannml
p rlge 1935) pp 245-290
WELFARE ECONOMICS 215
they relate to the planning of production under collectivism. By
avoiding all mention of utility and indeed without introducing even
the notion of indifference curves, Barone was able to break new
ground along lines which have in recent years become associated
with the economic theory of index numbers.
Unlike most of the writers discussed above, Barone is unsatised
with the statement that free competition maximizes product, or
sums of product, which can then be distributed in any given
fashion. Heterogeneous products cannot be added. Further
more, leisure may be preferred to the maximization of output. It
is signicant that those writers who do not explicitly introduce the
equations of general equilibrium should gloss over the denition of
product which is allegedly maximized. Thus, VVicksell con
nes his demonstration to a case where the same product can be
provided by different sources, and only in this case shows that
various marginal conditions are optimal. The same is true of the
very excellent treatment by Knight, in which the movement of
goods over alternative roads is analyzed for optimum conditions.15
It is remarkable that Professor Pigou, who reaches substantially
correct conclusions, never squarely meets the problem of the deni
tion of social product. His index number discussion represents an
important contribution in its own right, but it is offered at best as
an approximate criterion or indicator of changes in individual and
social welfare. He would not seriously suggest that the thing to
be maximized is the money value of output deated by an ideal
index of prices. Nor will the more exact limits of index number
theory such as are discussed in chapter vi avail.
Barone proposes to add different products after they have been
weighted by their respective prices; it is usually taken to be con
venient to express these prices as ratios to the numeraire good, a.
For Barone, productive services can be treated simply as alge
braically negative goods and services. Thus, decisions between
more or less work can be included in his welfare system. Then if
the sum total of each good consumed by all individuals together is
written as
A=+w+~o
B=w+m+~- @
etc.,
Wicksell, op. cit., p. 140, passim.
Knight, op. cit., p. 219.
216 FOUNDATIONS OF ECONOIIHC ANALYSIS
1nd recalling that
6U _ Uaa + Uob + (7)
we can write Pareto 5 equation (S) above m the equivalent form
1(60+ 6a+ ) + mob! + W + )+
=M+mw+ @
In passmg from (4) to (8) use rs made of the fact that the ratios of
marginal utllltleSof two goods for each Inleldual are equal to their
price ratios Barone himself does not use this terminology but
no doubt he would haxe to If the connection wrth Pareto were to
be shown
The expressron in (8) can be regarded as the variation in the
followmg expressmn \\ hen pnces are regarded as constants
=A+pbB+ (9)
If productive servrceswere known to be constant so that they could
be neglected (9) would equal (except for dynamic factors mvolvmg
capital which we can ignore) money value of national product If
all productive factors are included it Willrepresent the net differ
once bet een value of consumers goods and the return to produc
tive serVices Under many assumptions this quantity must be
zero under the full conditions of perfect competition
Barons shows that perfect competition maximizes this expres
sxon prrces being taken as xed parameters 1e any variation from
't condition of price equal to minimum cost must make <I> as gwen
m (8) negative Thus if we are at conditions other than perfect
competition With (8) not equal to zero for all posszble variations
it 111be possrble to spec1fy a movement which Will make 54p031
tlve But we can think of 64>as being made up of the sum of a
Similar expresswn referring to each indiVidual
54 = W + 640+ = (601 + Pbbl + )
+ (sa2 + p,,b+ ) + (10)
If for any movement the total 64>is pOSlthe it 15not necessary
that each and every one of the mdn idual 5 ap be posrtive but it
5 necessary that those which are posrtrve should outweight those
wlnch are negative Thus those who are hurt could be compcn
WELFARE ECONOMICS 217
sated by those who are helped, and there would still be a net gain
left to be parceled out among the individuals.
This is essentially the gist of the Barone argument. The one
point which will occur to the critical reader is the fact that arbi
trary prices are assumed in evaluating the expression to be maxim
ized. Which prices are to be used? Barone employs the prices
which prevailed before a contemplated break with the conditions
of competition are made, and this sufces if one merely wishes to
demonstrate that not all individuals can be improved by any
departure from competition.3
Unlike Pareto, Barone satises himself with deriving optimal
production conditions without going into the fact that under compe
tition no additional individual exchanges of xed amounts of goods
would be mutually protable. No doubt this oversight resulted
from his wish to avoid the use of indifference curves and utility,
but even without these constructions, by the use of the index num
ber notions which he pioneered, the enlarged conditions of exchange
could have been included. It is a tribute to this work that a third
of a century after it was written there is no better statement of the
problem in the English language to which the attention of students
may be turned.
8. The next writer who deserves our attention is A. P. Lerner,
who comparatively recently developed, presumably independently,
the Paretian conditions which show that the marginal equivalences
realized by perfect competition lead to an optimum of production
and exchange in the special senses discussed above.17 Actually, in
the eld of production his statement of the problem is slightly
different from that of Pareto and Barone. They showed that a
movement to conditions of perfect competition in the eld of pro
duction and cost could make everyone better off because they could
be given more of every good. But they still worked with indi
viduals. Even in a collectivized state where the existence of the
individual is not assumed, the Lerner formulation of the sense in
which output is optimal would still hold: the marginal equivalences
of competition are such as to give a maximum of any one product for
" Actually, Barone discusses varying prices in a passage which seems obscure to me.
0p. cit., p. 255.
" A. P. Lerner, The Concept of Monopoly and the Measure of Monopoly Power,"
Review of Economic Studies, I (1934), 157-175. Economic Theory 'and Socialist
Economy," Review of Economic Studies, II (1934), 51-61.
218 FOUNDATIONS OF ECONOMIC ANALYSIS
Because of diminishing marginal utility this is positive. but if we wish to have pi:
grCSSVC taxation, the elasticity of income after taxes against income before taxes must e
less than one. But
X d(Xt) _ XU'(X)
Xz dX _ (Xt)U'(X-t)
which is less than one if, and only if, the elasticity of the marginal utility curve is 155
than unity. Thus, for Bernoulli's law of utility, equality of sacrice would imply pro- _
,..
portional rather than progresive taxation.
228 FOUNDATIONS OF ECONOMIC ANALYSIS
quite a problem eVen for a man with denite opinions and great
preoccupation with value judgments If, however, he takes refuge
in assumption (5) that individual preferences are to count, the
individual can be left to decide for himself how he will spend his
money at given prices. Our ethical observer need only decide
then what his preferences are as between the given levels of satis
faction of different individuals
It might be thought that our ethical observer, even if the indi
\.iduals themselves had no unique cardinal indexes of utility, would
have to nd cardinal indicators But this would be quite incorrect
Of course, if utilities are to be added, one would have to catch hold
of them rst, but there is no need to add utilities The cardinal
utilities enter into the W function as independent variables if as
sumption (5) is made But the W function is itself only ordinally
determinable so that there are an innity of equally good indicators
of it which n be used. Thus, if one of these is written as
W = F(U. U. ). (13)
and if we were to change from one set of cardinal indexes of in
dividual utility to another set (V, l, ), we should simply
change the form of the function F so as to leave all social decisions
invariant Thus, let us move from one conguration of goods
going to the different individuals to another conguration which
leaves W unchanged or which xs socially indifferent. Then no
redenition of the Us or of F can change this fact the social in
differenceloci are independent of cardinal numbering And in this
terminology the signicance of assumption (5) that individual
tastes shall count is contained in the requirement that the social
indifference slopes between two goods gomg to the same mdmdual
are exactly the same as the individuals indifference ratios As
sumption (6) adds that these curves are unaffected by changes in
the goods going to other individuals
" Even If one Wishedto add ut1lnt1es,rtwould seem silly from any ethical wen pomt
to have ones opuuons as to correct taxation Inuenced by how consumers spend their
income on goods Yet that Is what the recent attempts to measure marginal utility
must amplyll' they have any pretensions to relevance for policy R Fnsch.New limited!
of Measuring Marginal Uhluy (Tbmgen, 1932) I Fusher, A Statistical Method for
Measuring'Margmal Uuhty' and Testing the justice of a Progressive Income Tax."
Economic Essays Contribzdrd m Honor of John Bates Clark (New York, 1927)
WELFARE ECONOMICS 229
MATHEMATICAL ANALYSIS
a_X.
69]: _ _ vm; _ TX; _
E -
EE r (ale, l 123, (21)
av: aime
a av, Ti" .
a1=--o=a)n=7- (],r=1,---,m) (22)
Ov,-1 avr)!
It will be noted that only ratios of marginal utilities for the same
individuals are ever involved. Thus, there is no need for nu
merical utility even for a single individual and no need to compare
the utilities of different individuals.
Graphically the equilibrium may be represented by the same
box-type diagram as was mentioned in the last section. The di
mensions of the box represent the xed totals of the commodities
in existence. Any interior point oriented with respect to the
lower left-hand corner represents the amount of consumption of
the rst individual, and the contour lines are now to be interpreted
as indifference curves. Referred to the upper right-hand corner,
a point represents the second individual's consumption, and his
contour lines may be superimposed on the other. Specifying the
Equilibrium conditions regarding the factors of PdeUCOneed not be indicated
separately if it is remembered that they may be treated as negative commod
238 FOUNDATIONS OF ECONOMIC ANALISIS
utxlxty of the second 1nd1v1dual,we maxxrmze that of the rst by
movmg along the gwen mdl'erence line of the second 1nd1v1dual
untrl ne reach (at a pomt of tangency) the hlghest 1nd1flerence
curve of the rst Because the ongmal specncatnonof the second
man's utxhty was arbitrary, the nal equrhbnum xsalso arbitrary
and not umque The locus of all such arbltrary pomts 15,of
course, the famrhar Edgeworthran contract curve and represents
the set of ponnts satrsfymg necessary exchange condmons
After the dtscussxon of the prewous section 1t Willnot be neces
sary to reiterate at length the contention that the equahty of the
xanous ratxos can be expressed thhout Lagrangean multrphers
and wrthout prlce ratlos, moreover, m by no means 1mp0551ble
cases, even playing the game of competmon would lead away from
rather than to ard the correct optimal posmon As before, thus
lS because the secondary condrtlons for a constrained maximum
are not the same as the secondary condxtxonsnecessary hterally to
brmg the Lagrangean expressron to an extremum
It Willbe noted that from any pomt off the contract curve there
exnsts a movement toward rt uhlch would be benecxal to both
Indrvrduals Thls 15not the same thrng as to say, wnth Edgeworth,
that exchange wl fact necessarily cease somewhere on the con
tract curve, for m many types of brlateraI monopoly a nal equr
hbrmm may be reached off the contract curve Nor 131t the same
as to say that pomts on the contract curve are better than pomts
off the contract curve Later I shall dlSCLISS the correct formula
tron of the sngmcance of thIS condltlon
But rst rt ISn ell to write down the full rst order condltlons
of equlhbnum wluch must hold 11'we are to have an optimum of
both productlon and exchange, that ls, If we combme w1th the
condmons of thlS sectnon the condttrons of the last sectlon Then
(1) ne must have a common margmal rate of zndzerence betweenany
mo goods for etery matvtdual, this common andz'erence 111110must.
moreozer, be equal to the ratzo at wind: one of these goods can be
transformed mto the other on a productzon sense. the transformatton
to come about as the result of transferring any resource from one
good 3 productzon to the others (2) We must have for all mdwzduals
a common who of mdzerence between supplymg more of any factOT
' praductzon and enjoymg more consumptzon of a gwen g00d: th
WELFARE ECONOMICS 239
common ratio must be equal to the rate at which supplying more of
that factor results in greater production of the good in question.
Mathematically, as Lange has shown,
:} Un T_Jn
U3} = . . . = sz : Tx_ (ii,7.
: 1, . .., )
Uk1=___=_ULJ__7_ (i=1,---,1z) (26)
UZ,-1 U,: " am,, _ Tx (k = 1, . . ., m)
-..._6__= <=L~un>
st x. GX,- (J = 1, - - -, m) (30)
6W _.'._____-._-._...
3W 6W 6W6X1 ___
6W 6X" .
246 FOUNDATIONS OF ECONOMIC ANALlSIS
Not only are these equatxons very snmllar to the famous Note
XIV whrch summanzes much of u hat xs best Marshall s Prm
czples, but they are cast almost exactly m the form followed by
Professor Plgou m hls Economzcsof Welfare In words, two condr
tlons are Implied rst, the margmal soczal utzlzty (dzsutzhty) of the
some good (servzce) must be equal for each mdwzdual, second, each
factor of production must be dwzded among dt'erent posszble uses so
that tts mdzrect. derwed margmal saczal utility must be the same m
every me and equal to mmargmol soczol dtsutzlzty
We may leave to the reader the task of showmg the necessary
modtcatrons 1f the supply of a factor IS Inelastic, If a factor of
productzon 15not mdlerent between dlfferent uses If there lS]omt
productlon, nf there are external technological economxes or dlS
economies (so that the production functton of a good contams m
xt factors of productxon devoted to other uses), etc By dmdmg
the above equattons through by any smgle partial denvatxve. they
may be thrown mto a form which ISIndependent of the part1cular
cardmal representatlon of W
Whlle the Lerner productlon condltlons are contained m the
above set of equatlons the exchange condltnonsare not However,
If we make the 1nd1v1duahst1cassumptlons ve and 81xabove so
that the welfare functxon takes on the spectal form gwen m equa
tron (13). then m Vlr-tueof the 1dent1t1es
6W W U'
xJ= _U'x: (31)
%fj
= ma). (s)
dx.
da=_ :
T j: AH ' @
where the subscripts indicate partial differentiation,
fx f2: ... fn
A=fff =Ife'l.
f 1" f 2" f n"
and Ai;is the cofactor of the element in the jth row and ith column
of A.
Unless some a prion" restrictions are placed upon the nature of
the elements involved in these determinants, no useful theorems
can be derived. Each unknown derivative depends upon an
n(n + 1) innity of possible values. If the various determinants
were expanded out, a sum of n! terms would appear in the denomi
nator and in the numerator. Regarded simply as chance drawings
taken at random from a hat, the probability that the signsof these
would all agree would go rapidly to zero as the number of variables
increased. Fortunately, as will be shown, the analysis of the
stability of equilibrium will aid in evaluating these complicated
expressions.
If for a given value of a = a1, there exists a solution (xr, -- -,x..'). and if the
matrix [af/ex,] is ol' rank n in a neighborhood of (x), then by the implicit-function
theorem equations (2) represent single-valued continuously differentiable functions in
a sufciently small neighborhood of (ax. x).
260 FOUNDATIONS OF ECONOMIC ANALYSIS
In the Simpleexample of supply and demand alluded to above
our unknowns are (p. q), and our equnhbnum system can be wntten
gD(pa)=0 (D.>00,<0)
qS(p)=0 (5)
where a ts a parameter of shtft representing taste and D,, 1%
usually assumed to be less than zero Also
31_ . __1__
$ Da Spo __ Dpo v (6)
(19:
= D o __.P_. (7)
FIGURE 2 FIGURE 3
@
(dxl) 513)II.
> (de (35)
The equilibrium is approached asymptotically. .
If both curves are negatively inclined, stability requires
(dx1)1<
1322 dx] II .
(!) (36 )
Clearly, the general condition when the curves are of like sign can
be written
Ea ! a (37)
dxl I dxl II
the ks depend upon the matrix a and upon the initial conditions.15
As before, stability requires RO,)< 0.
By a well-known theorem of Hermitian matrices, in the sym
metrical case all the roots are necessarily real. If the equilibrium
is to be stable, they must all be negative. According to a classical
theorem, this is possible if and only if a is the matrix of a negative
denite quadratic form; i.e., only if all principal minors alternate
in sign as follows:
Gif Gif" Gaz
au. 0 au..0
|a.-.-|< 0; . .o . .0
> 0; cf.-" 013 u" < 0.
a: 3 au
Ghia ah,- au"
ijki MD
combination of goods and services would, strictly interpreted. lead t an absurdity; it
would imply expenditure proportionality and, hence, zero consumption of every good and
zero offering of every service! For the general demand or supply function we need nt
expect a canceling off of income e'ects" since individuals usually face rms in consump
tion and factor markets.
" If the roots are not distinct, polynomial terms of the form le. e, ' ' "
appear where (: + 1) is the multiplicity of a repeated root. In any se. the Problem
of stability depends only upon the xs and is unaffected by such multipliers since the
exponential always governs the asymptotic behavior of the solution whenever dampening
does occur.
272 FOUNDATIONS OF ECONOMIC ANALYSIS
Any ['3th of the form
l 0 0
0 a)! 1&0
0 au du (48)
a "o a "o a tl.
0: o a. an"
where q 15the column of q.'s a the square matrlx a,, and where H
need not now be a posnt1ve dlagonal matrur It would seem
reasonable at rst to requnreH to be posmve demte, or at least to
be posmve quasn demte as the latter expressnon was dened m
chapter vx But If we apply the contragredxent transformation c
as gwen m equation (47) of chapter V1,
9 ""' Cg, : (:_lg
? = ? ? = C'? (52)
" The followmgIllustrations bear this out The system
1h = 291 + 4P:
P: = " PI + ?:
possesses stability of the rst kmd but rs neither perfectly nor Imperfectly stable
The system
p: = Pa "' Pa
11: = 291 + P:
Is Imperfectly stable but departs ever further from equlhbnum
"O Lange Abstract of The Stabxlty of Economic Ethbnum Ecanomemca
X (1942) 176177Puce Henbddy and Employment (Bloommgton Indzana Pru'lcmla
Press 1944) append.
STABILITY AND COMPARATIVE STATICS 275
then it can be veried that
13=Hq+---=c'Hq+---=Ha+~-= (cHac")13+---.(53)
It follows that our restrictions cannot be placed upon H alone,
but only upon the product of H and a; that is, the results depend
upon the statical matrix a and upon the dynamical reaction-speed
matrix H. It is sufcient for stability that in some coordinate
system H be positive quasi-denite and a negative quasi-denite,
but these conditions are by no means necessary.
Before leaving the problem of stable multiple markets, I should
like to sketch the e'ect of the introduction of stocks and its rele
vance to stability of the second kind. Let price fall not when cur
rent supply exceeds current demand, but when existing stocks (ac
cumulated over time from the divergenceof current production and
consumption) exceed an equilibrium amount. Then for proper time
units and equal speeds of adjustment
where la.-5- Mia-I= 0, and where for unrepeated roots the k's
and the hsare constant depending upon initial conditions. Clearly
the motion will be explosive and undamped unless w/X;are all pure
imaginary numbers; i.e., unless ,-is real and negative.
If the system is symmetrical, this clearly leads to the same
conditions as those for stability of the rst kind. If not sym
metrical, the substitution of second derivatives everywhere for rst
derivatives (through the hypothesis of dependence upon accumu
lated stocks rather than instantaneous ows) implies more ngld
conditions upon the coefcients to insure stability of the second
kind than were previously required to insure stability of the rst
276 POUNDATIONS 01 LCONOMIC ANALYSIS
kind. This is of course because of the requirement that the roots
be real as well as negative.20
ANALYSIS OF THE KEYNESIAN SYSTEM
.=anl:am+("h)]
a o ( ap>0
> 0)
if stable for B > O, a = 0, and also {or B = 0, a > 0. it can perhaps be proved to bi
stable for all Intermediate cases
" J E Meade, A Simplied Model of Mr Keynes' System." RMN of Economy
swam. IV (1937), 93-107. J R H1cks,Mr Keynes and the 'CIaSSICS'.A Suggest
Interpretation," Econometma, V (1937), 147159.Oskar Lange. The Rate of Interest
md the Optimum Propensuy to Consume," Economica. V (1938). 12-32
STABILITY AND COMPARATIVE STATICS 277
an upward shift in the propensity-to-consume schedule; similarly
as the parameter ;?increases, the marginal-efciency schedule shifts
upward.
We have three relations to determine the three unknowns in
terms of three parameters, viz.:
i = f(a, B, M).
Y = Y(a,, M). (59)
I = I(a, B,M).
As explained in the rst section of this part, the usefulness of the
Keynesian equilibrium system lies in the light it throws upon the
way our unknowns will change as a result of changes in data.
More specically, what are the signs of
d; a il.
da, da' da'
ii LY 115
615, dB' d'
di dY (il
5% ar m?
Differentiating totally with respect to our parameters and evaluat
ing the resulting linear equations, we nd
_di_LY Q__L_" _FLi-FiLY
dcr A ' docA' da A '
__LY __L_.- =(1CY)L'+CL}"
dB _ A ' dA' (16 A (60)
_1CY-FY _Fi+ci
dM_ A ' dM_ A '
il; __Fy(F.+ Ci) + (1 "' Cy "- Fr)]:
dM A
where
C.- Cyl 1
A= F.- Fr _.1 =LY(F.+C.-)+L.-(1CrFr)- (61)
L.- Ly 0
On the basis of a priori, intuitive, empirical experience the
followingassumptions are usually made:
Cr>0. Fy>0, F.-<0, Ly>0, L.-<0, (62)
278 FOUNDATIONS OF bCONOMIC ANALYSIS
whtle >
C . *2-0
and xsusually assumed 1n modern discussmns to be of minor quan
txtatwe Importance
In order to evaluate our nme denvatlves we must be able to
detennme unambiguously the sxgnsof all numerators as well as the
common denommator, A A consnsts of ve terms, two of whlch
are of posntwe sxg'n,two of negatxve SIgn, and one ambrguous On
the bas:s of deductlve analySISalong stnctly statncal lines nothmg
can be Inferred concermng 1ts sign Moreover, even if the sxgnof
A were determmed, all but four of the mne would be found to have
numerators of mdetermmable Sign
Thls Is a typlcal case If we are to derrve useful theorems ue
must clearly proceed to a consnderatron of a more general dynamlc
system whlch meludes the statlonary Keynesxan analy51sas a spe
ela] case Thls can be done m a vanety of alternatwe ways I
shall consxder two, the rst of which xs based upon a dxerentlal
system and ylelds quite demte results
Case 1 Let us assume as before that the second and thlld
relations of margmal efcncy and llQUldltypreference work them
selves out m so short a tlme that they can be regarded as holdmg
mstantaneously Let us assume, however, that I now represents
Intended Investment, and this magnitude equals savmgs mvest
ment only m ethbnum, 1e , when all the vanables take on sta
tlonary values If. however, because of some change, consumption
(say) should suddenly Increase, national mcome not havmg a
chance to change, actual savmgs 1nvestment would fall short ol
mtended mvestrnent because of Inventory reductlon, etc Con
sequently, Income would tend to use Snmllarlyan excessof actual
savxngs Investment over Intended 1nvestment would tend to make
Income fall Mathematically, this hypothesm may be stated as
follows the rate of change of mcomc zs proportzonal to the dzcrence
between mtended samngsmvestmentand actual savmgs mvestmen!
The dxscusswnhere ISunrelated to the controversy over the equal
Ity of sangs and Investment despite possnbleappearances to the
contrary The supercnal resemblance between my formulatlon
and the Robertsoman Identltres whereby the dlerence between
Investment and savmgs xsthe tlme dlerence of Income should not
m-slead the careful reader
STABILITY AND COMPARATIVE STATICS 279
Equations (56), (57), and (58) are replaced by the dynamical
ones:
Y=I-|IY-C(i.Y)-al (63)
0 = F(z', Y) I+B, (64)
0 = L(i, Y) -- M. (65)
The solution of these is of the form:
Y = Y"+ ale,
i = io + 028, (65)
I = I + age,
where
C,- Cy 1 A 1
A0) = Fi Fr - 1 = A + XL.-= 0. (67)
Li LY 0
The equilibrium is stable only if .
A
= f.- < O. (68)
But L.- < 0; therefore,
A < O (69)
unambiguously.
This establishes four theorems: an increased marginal efciency
of capital will (1) raise interest rates and (2) raise income; an
increased propensity to consume will (3) raise interest rates and
(4) raise income. But how will the creation of new money affect
interest rates? This can be answered by considering more string
ent stability conditions. Let us suppose that the interest rate
were kept constant (say) by appropriate central bank action.
This assumption is equivalent to dropping the liquidity preference
equation (65) and treating i as a constant in the remaining equa
tions. If the equilibrium is stable under these conditions, we
must have
1
Fy
1=0=(1CyFy)+>\, (70)
- 1
OI'
>\=(1FY_CY)>O. (71)
This leads to another important theorem: (5) the marginal pro
pensity to consume plus the marginal propensity to invest cannot
280 FOUNDATIONS OF LCONOMIC ANALISIS
exceed mzzty or the 33mm mil be unstable (as of a xed Interest
rate) It also tells us (6) that an merease 111the amount of
money must, cetens parzbus 10er Interest rates
We are left n 1th four ambigumes of szgn Tho of them depend
upon the fact that savmgs may \ary m any dlrcctlon 11threspect
to a change m Interest rates If we assume that normally sawngs
out of a gwen mcome mcrease wnth the mtcrest rate, or. nfthey do
decrease do so not so much as does xmestment, then three more
theorems become true 'un Increase In the amount of money (7)
mcreases Income and (8) Increases nmestment (9) an increase
the margmal cfmency schedule Increases Investment There re
mains a nal amblguous term 'h'lt ISthe effect upon Investment
of an mcreased propensnty to consume? Thus is seen to be essen
txallyambiguous depending upon the quantltatu e strengths of the
llQllldlty preference slopes and the marginal cfcrency slopes As
Income Increases money becomes tight because of the need for
nancmg more transactrons Thls tends to depress Investment
As an oset the Increase m Income tends to Increase Investment
through the margmal propenSIty to lmest thch effect wrllbe
the stronger cannot be decndedon (: pnon grounds
I have prepared a 3 X 3 classncatron Indicating the s1gns 01
the mne terms All but four ham demte sxg'ns Of these four
one ISessentially ambiguous as Indicated by a questxon mark The
remammg three show under question marks thelr normal pre
sumptlve sxgns
z Y I
Increase m propensnty to consume a + + ?
Increase 1n margmal efcxencyof capital B + + :3
Increasem amount of money M - l l
A + L.-Cy < 1.
m = lLC" (86)
1FY<0<CY<(Fy-1)-%(F.-+C.-), (88)
the equilibrium will be stable, but the signs of our 3 X 3 table now
are as follows:
Y
a
[3
In words, the only theorem which remains true under all circum
stances is that an increase in the amount of money must lower
interest rates if the equilibrium is stable.
This example illustrates the additional complexities which sys
tems based upon di'erence equations involve. Later some of the
reasons for this will be explained.
The examples here adduced serve, I hope, to illustrate the light
which dynamical analysis sheds upon comparative statics. Prob
lems in theory and business cycles of any complexity will almost
surely require similar analytic treatment if useful and meaningful
theorems are to be derived.
CHAPTER X
Bm+wm=a m
where
mw=w+J:mw-mwa G=L~uw(M)
283 fOUNDATIONS OF ECONOMIC ANALISH'
are also of mterest Unfortunately the theory of nonlinear
tegral equatxons xsonly fragmentary
men mm Newman SYSTEMS
Up untnl now most economxsts have concerned themselves wnth
lmcar systems not because of any behef that the facts were so
snnple but rather because of the mathematzcal difculties Involved
m nonlinear systems Thls rs understandable and excusable smce
all thought advances m small steps Nevertheless from the stand
poxntof a study of mdustnal uctuations this can be a rather sen
outs llmrtatlon Thus m a linear system the amplltude of uctua
non depends upon the mmal dnsplacement no Intrmszc amphtudc
-asbetween full employment and zero employmentns mvolved
The attempt to Introduce such a xed amphtude m a lmear system
by the devrce ol determmmg coefcxents so that there Will be
nentherdampened nor exploswe solutrons appears to be nusdzrected
As pomted out In the prevnous chapter page 268 the stochastxcal
dispersmn of the system Increases 1ndemtely Related to this 15
the fact that thusdoes not yield a umque amplltude but one whlch
depends lxnearly upon mmal condrtnons
If we msnst that a system be hnear and that It do not Involve
tlme exphcntly then {or dlfferentxal and dxfference systems we are
restncted to the case ol constant coefczents Thxs type 15mathe
matlcally sxmple and exact solutnonsare known But a hrgh puce
as paid for thrs snmphcnty In terms of specral assumptlons Wthh
must be made
%=X=f(X)=Ao+A1X+AzX+ (15)
wheref ISanalytic and expresslble as a power serles This equation
presents no difcultres of solutnon smce by a smgle quadrature 1t
LINEAR AND NONLINEAR STABILITY 289
may be solved; namely,
X dX
llg=j; =F(X). (16)
Let us suppose a simple" stationary solution X = X exists
so that
f(X) = 0.
f'(X) # 0_ (17)
Then the transformation of variables
x = X X" (18)
transforms (15) into the form
2 f(X)x
:i'=f(x+X) = Li' =0+a1x+agx2+ ---, (19)
where a constant term is no longer involved, and al does not vanish.
Then we can assert:
THEOREMI:A formal solution of the diferential equation (19) can
bewritten as an innite power series in the solution of the simple linear
system
:i: = (11x, (20)
07
C. = M(Cnx. . Cl)
K _. _1__ _ 1. (32)
_ x(0)
For K = 0 we have the stationary solution
x(t) E 1, (33)
292 F0 UNDATIONS OF ECONOMIC ANALYSIS
and for K = we have the statxonary solutron
xe) % o (34)
It can be easnlyvened that the latter stationary level :s un
stable, while the statlonary level of umty rs approached asymp.
tottcally by all adjacent motlons We may classxfy all posmble
mltml conditions as follows
+lx(0)._"E, 1<K0, } __
o<x(0)-==1. ogK<+ oo. '.LE"() 1' (35
x(0) :O, K: oo, xEEO, )
oo<x(0)<0 -oo<I&<l, lrmx(t)=oo
Let us now apply our expansxon theorem to thns problem
Expandmg around the zero cqmllbrlum pomt. and determlntng the
c coeiczents we readlly nd
x = ae - {ael + {ad}:
lael + = (~1)1[ae'} (36)
1
: Xlt) l+KQ
I
, 2.0
I
l
: -l.5
I "VK
K-O II
I
___
4K4
I 5//_
I
I
I0
.
K=
' L
. . : l + : .
-4 ..3 ..2 -I O I 2 3 4
FIGURE 4
F(X) = IXKXMX
Then the lollowmg sums up the results achlcved
THEOREMIV (a) If F(X) a'ords a relatzte marzmum to F.
then X" as a stationary solutton of the dz'erentzal equatzon and
possesses stabzlzty m the small, and comersely
(b) If F(X) a'ords a relatne mzmmum to F then X 1s an
unstable egmltbrmm level
(6) If X as a stationary talue of PM ) uhtch es not an ettremum,
then the system possesses one szded stabzlzty mstabzlzty Alterna
twely af F assumes a statzonary talue and F = f assumes an er
tremumwine, the ethbmcm :: stable unstable
(d) If F'(X ) tanzshestdenttcally.the ethbmcm zs neutral
Thls possnbllatyof lmkmg up the problem of stabxhty thh a
statical maxnmum problem [5 but one spccml aspect of the Corre
spandence Prmczple and one to h1ch ue shall have occasxon to
reier again
AN EXAMPLE OF OM: SIDED STABILITY-INSTABILIT}
VIALTHUSIANAND Oprmuu POPULATION THEORIES
The meamng of one Sided stabxllty Instability may not be
tmtnvely obvmus fortunately a Simple well known economlc ex
ple may be used to Illustrate at Accordmg to Malthus popula
LINEAR AND NONLINEAR STABILITI' 297
tion would increase, decrease, or remain stationary depending upon
the per capita level of subsistence (real income, food, etc.). Let
X = total population, S = per capita real income. Then the
percentage rate of growth of the population is an increasing func
tion of the level of subsistence, passing from negative values to
positive values at some minimum level of subsistence," 5.
Mathematically,
% = MS). (51)
where
MS) > 0. M5) = 0
But the level of income (production) depends itself upon the
level of population (labor) as of given capital, land, and technology.
Moreover, Malthus implicitly and explicitly assumed the law of
diminishing (per capita) returns. Thus,
5 = (X), '(X) < 0- (52)
This last relationship enables us to eliminate S as a variable and to
express the rate of growth of the population in terms of itself.
I .
u
:| 1L
t ll
mu
can:
't!
REM.
r LE
tom. retour On
FIGURE 5
FIGURE 6
Nonsingular A
.Neutral G
Nonsingular F
LNeutral J, T
306 FOUNDATIONS OF ECONOMIC ANALYSIS
We dzscrxmmateanalytlcally between the szxteen cases depend
mg upon the rst and higher denvatrves of f(X ) The rst sub
dlvrsron between monotonie and oscrllatory hmges upon whether
f(X ) rs posrtrve or negatrve Wnthm the monotomc classxcatnon
we haxe rst order stabnhty 1ff rs less than umty and rst order
mstabahty nlf asgreater than unity If j' equals unity, we have
rst order neutrality and must proceed to hrgher derivatives If
all these denvatlves vamsh we have true neutrality m the small
If the rst nonvamshtng denvatwe 18odd and posrtrve, we have
high order stabllrty. llodd and negatwe, ue have high order m
stabrhty If the rst nonvamshtng higher derrvatne as even, we
have one srded stabrhty 1nstab1hty
W1thm the oscrllatory classrcatlon things are even more com
phcated If )" xs less than umty m absolute value, we have rst
order stablhty, ll greater than unlty m absolute value, we hate
rst order Instabllrty ll' f = -1. then we must proceed to
higher denvatxves Should all of these vanrsh we have neutral
osc111atoryequthbnum (1) More generally, If all odd hlgher
denvatwes vanish, the equlhbnum IS neutral and oscrllatory as
ln T When f = 1 and the rst nonvamshmg denvattve 15
odd and posmve, the ethbnum possesseshigh order oscxllatory
stabllrty ll odd and negatwe, the equlhbrxum possesses hnghorder
oscxllatoryxnstabxhty When]" = - Land the rst nomamshmg
derlvattve ISeven then we go on to consrder the rst nonvamshmg
odd derxvatlve As m the prevnous cases, we have stablhty or
mstabnhty depending upon whether thrs ISnegatxve or posntlve
Fmally, we are confronted thh the smgular cases whenf' = O
Il the next nonvantshmg denvatwe rs odd and posture, we have
monotomc stabrhty. ll odd and negative, or If even and of any sagn,
we have stable oscnllatory eqmlrbrnum If all denvatrves vamsh,
the equthbrmm ISperfectly stable as mdxcated m the rst sectton of
thls chapteravhen drsplaced the system returns mstantly ' to equx
hbnum and does not merely approach equrhbnum asymptotically
A few remarks may be m order concernmg the qualttatlve be
havwr of a rstordersystem from any 1n1t1alcondltnon It could
no doubt be shown that 1t must do one of the following (a) go off
to mmty. (b) approach an equlhbnum level, or (c) approach 1
penodlc motion of some mte penod If lt :s revers1ble, that IS
lf f(X) ISnot only smgle valued but adrmts of a smgle valued m
---_.-- _--. -uv-.u4vuan4\ U (11.11.1415
1 du,
verse, then the only periodic motion possible under heading (c) is
one of period two. It is not appr0priate to investigate here the
meaning of the stability of motions more general than stationary
equilibrium levels. When this is done, it will be found no doubt
that there exist valid separation" theorems concerning the order
ing of stable and unstable periodic motions. The necessity because
of continuity (Rolles theorem, etc.) of duplicating certain equi
librium points in the above diagram provides a hint at such
relationships.
ANALYTIC SOLUTION
{Xa(72) _ Xgo}dTIdT2 + - . .,
where K.,, K represent rst and second functional derivatives
respectively. The linear system
X ll ix: + {hd,
X 1Xs'l'll' {mc}, .
J A Schumpeter Theory of Economu: Development chap :
FUNDAMENTALS OF DYNAMICAL THEORY 317
where in the rst case the random variable 12which may appear at
any instant of time is drawn from an unchanging universe; and
where in the second case the random variable m is drawn from a
universe dened differently at each instant of time."
In his recent book Value and Capital, Professor Hicks has
given an exceedingly simple denition of dynamics: 1 call Eco
nomic Statics those parts of economic theory where we do not
trouble about dating; Economic Dynamics those parts where every
quantity must be dated (p. 115).
In terms of the above six categories this denition is overly
general and insufciently precise. The second category consisting
of historically moving static equilibria would certainly require
dating of the variables; but it would not thereby become dynamic.
My objection is to his denition, not to his practice, for many of
the systems which he analyzes are in the strict sense dynamic.
CAUSAL SYSTEMS
% + x0) = :. (ss)
where the right hand member IS treated as a constant Clearly,
such a posrtlon ISgwen by
O+x=t (36)
But this lSpreusely the rst approximation x10) of equatlon (25).
Specrcally m thns case
dx
d;=0=zxl (37)
The Virtual path of the system after the freezmg of hlstorlcal
change ISonly one possrble solution For other mntxalcondrnons
at tlme : dl'erent solutlons Willbe dened In partlcular, let us
consxder lmtlal condmons u(i, a:), where u 18a true solutlon of the
lustoncal system (34) The corresponding Virtual path may be
written
:. & v, u(t, x)] (38)
For the specxcsystem dened by (34)
u(t, :) = : -- 1 + (a:+ De, (39)
aft, u(t, z)] = ( c + x + De +! (40)
FUNDAMENTALS OF D YNAMICAL THEOR Y 327
The actual solution of the historical system is seen to be an envelope
to the family of virtual solutions vl], u(i, i)], The virtual move
ment for short times approximates closely the true movement in
that they are mutually tangent. That is, if the boy weresuddenly
to stop walking, the pendulum would change its motion; but the
new motion would be tangential to the old, and for short intervals
they would not diverge much.
The virtual curves approach asymptotes or hypothetical sta
tionary equilibria. These equilibrium values plotted against time
form the curve of receding equilibria. This function of time is not
in general a solution of the true dynamical equations. The sta
bility of the hypothetical equilibrium can, however, be determined,
and, as will be seen later, bears some relation to the stability of the
actual motions.
3. I turn now briey to the normal values or equilibrium
values dened by Professor Frisch. This is dened by relaxing
a subset m of our n functional equations
. u" {"
328 FOUNDATIONS OI ECONOMIC ANALYSIS
in the barred variables, (51, - . i.). This denes the moving
equilibrium.
Professor Frisch is well aware that this denition does not
correspond in the stationary case to the concept of equilibrium; he
even thinks that the former will eventually replace the latter as
the tendency to formulate the economic reasoning in exact dynamic
mathematical terms gains ground. But this concept, he be
lieves, is what many modern writers have in mind.
Whether or not this last conjecture is true, I do not know, par
ticularly since I am unacquainted with the original Scandinavian
writings.27 But for the present purpose this is irrelevant; it is
only important to point out the differences between this concept
of a normal and the moving equilibrium for which Weare looking.
Some of the salient properties of the normal" movement are.
(1) It is in general not a solution of the original set. There
exists no set of initial conditions which would produce it in actuality.
(2) It is not unique, but depends upon as many parameters as
are necessary to specify initial conditions for the original Set of
equations This is because it depends upon the particular solution
out of an innity of posable solutionsinserted in the modied
set of equations f and g. (See the italicized sentence above.)
(3) Being dened by essentially static, albeit possibly historical
equations, there is no sense in asking questions concerning its
stability.
For all these reasons it is unsatisfactory as a representation of
moving equilibrium for our present purposes
[bid , p 102
" Professor Frisch illustrates this concept by referring to the Wicksellian relation
between the actual and natural rate of interest. In so rationalizing the arguments of
some neo-chksellians he is, I fear, over charitable, ascribing to them an undeserved
degree of sophistication Actually, Wieksell himself considered for the most part a
system of impliut dynamic relations not involvmg time explicnly. i e , Without historical
change (Witness. for example. his controversy with Professor D-iindson on the equi
librium requirement of stable prices in a system undergorng output expanSion because
of irreverSible technologin change ) ithin a nonhistoricnl framework it is possible
to construct dynamic models of the Wicltsellian system in which the natural rate of
interest represents the stationary equilibrium level of the system and not the solution
of a hypothetiml alternative system If this be granted, what is the corresponding
condition for a system in which elements of historical change have been introduced?
This brings us back to our original query.
Still another notion of morvmgequilibrium is suggested by an explosive exponential
FUNDAMENTALS OF DYNAMICAL THEORY 329
RESOLUTION OF THE PROBLEM
i.e., for any given positive e. however small, there exists a t such that
lx(t) xl < e, for t > 1".
But this is symmetrical in :(!) and x so that the equilibrium position can be said to
approach any motion, as well as vice versa. It easily follows that it"(t), any motion,
approaches :la), a particular motion selected arbitrarily.
33o FOUNDATIONS or ECONOMIC ANALYSIS
1nc1dencethat this problem could be studied by an exammatron of
the properties of a smgle specxalone It wrll perhaps be enlighten
mg to pomt out that even 111the nonhxstoncal case we can Just as
well concentrate upon any other smgle motxon than the equlhbnum
one For 1fthey all approach the equrhbnum motion then they
all approach each other And every motlon must approach any
one selected at random
Thxs suggests the followmg answer to our dllemma We shall
concern ourselves not wzth the stabzhty of a pamcular manon of a
hastortcat system winch can be gwen the prwzleged tztle of momng
egmhbnum but rather unth the stabsltty of each and ezery manon of
the system Thls WI" not preclude our deSIgnatmg for specnal at
tentlon partlcular mot10ns wrth specnalproperties (as for example
stnctly perrodlc motions steady motion m whlch some coordmate
Increases Imearly etc)
Thns suggests how I would answer the purely verbal problem
what processes shall be demgnated as egmlzbrzumprocesses? Once
we have strlpped the equrhbnum concept of normatlve and teleo
logncal connotatlons 1t does not much matter how we apply the
term The term may be applted to stattonary \ alues only From
thus narrow pomt of vrew a competntlve Industry would be m CQUI
hbnum only after all long run condltlons had been satlsed entry
of rms correct output for each rm etc
For other purposes we mlght apply the term to those mter
mediate short run sntuatlons m whlch each rm IS producmg at
pnce equal to margmal cost (and above average vanable cost)
even nf price dxverges from average cost so that the number of
rms ISchangmg or xsabout to change It rmght be reserved for
Situations of the last descnbed type where the further condrtron IS
reahzed that the number of rms rs changmg at a causally deter
mmed rate Stxllothers may be wrllmg to apply It to momentary
condrtlons of supply and demand 1nclud1ngvarlatlons m speculative
stocks of the good m questlon
I, myself nd lt convement to vnsuahzeethbnum processes
of qunte dxfferent speed some very slow compared to others
Withm each long run there ls a shorter run and thhm each shorter
run there ISa stxllshorter run and so forth m an mmte regresswn
For analytlc purposes 1t lSoften convenrent to treat slow processes
as data and concentrate upon the processes of mterest For 6!
FUNDAMENTALS OF DYNAMICAL THEORY 331
ample, in a short run study of the level of investment, income, and
employment, it is often convenient to assume that the stock of
capital is perfectly or sensibly xed. Of course, the stock of capital
from a longer run point of view is simply the cumulation of net
investment, and the reciprocal inuence between capital and the
other variables of the system is worthy of study for its own sake,
both with respect to a hypothetical nal equilibrium and the simple
course of growth of the system over time.
So to speak, we are able by ceteris paribas assumptions to dis
regard the changes in variables subject to motions much slower"
than the ones under consideration; this is nothing but the per
turbation technique of classical mechanics. At the same time we
are able to abstract from the behavior of processes much faster
than the ones under consideration, either by the assumption that
they are rapidly damped and can be supposed to have worked out
their effects, or by inclusion of them in the dynamical equations
(derivatives, dierences, etc.) which determine the behavior of the
system out of equilibrium.
The rst of the above mentioned alternatives constitutes the
justication for the use of comparative statics rather than explicit
dynamics. ~ If one can be sure that the system is stable and strongly
damped, there is no great harm in neglecting to analyze the exact
path from one equilibrium to another, and in taking refuge in a
mutatis mutandz'sassumption. Of course, if one chooses to neglect
certain dynamic processes, one may still retain others; e.g., in
studying capital formation over two decades I' may choose to
neglect inventory uctuations, but still may retain the acceleration
principle in its secular aspects.
Under the second' alternative where shorter run processes are
contained in (say) the differential equations of the system, it is to
be understood that these differential equations do not necessarily
hold exactly at each instant of time. There may well be a still
shorter run theory which explains how still higher differential equa
tions lead to (rapidly) damped approaches to the postulated differ
ential equation relations. And so forth in endless regression.
It may be argued that so general a connotation is at variance
with traditional usage of the word equilibrium. Is it not straining
language to think of a cannon ball as being at equilibrium, not
only after it has fallen to the ground at rest, but also at every point
332 FOUNDATIONSOF sconowc ANALYSIS
m Its lght, when It ISon Its mean trajectory as well as m 1ts pre
cessxonaround this path? Perhaps such termmology may occa
sronallylead to confuswn, however, thh carefully stated quahca
trons It may be convenient
To examme the stability of any motion [u,(t) , m)] sunply
substxtute mto the functional equatlons
x.(t) = 24.0) + m0) (43)
where the above equation ISSimply '1denition for the :7'5 There
result n functnonal equatlons m the r;s smce the u s are gwen func
trons Because the u's constitute a solutlon of the ongmal system
there emsts a solution m the n s of the form (0 0, ,0) The
ongmal solution w1ll be stable nf for all poss:ble :mtlal conditions
hm n (t) = 0 (1 = 1 n) (44)
lou
fEufr). :] = 0 (46)
[Qmma=ma mm
and
Emva=na m)
the general linear differential and difference equations can easily
be written as linear integral equations. Thus,
% + x(t) = 0 (50)
takes the form
_ K(t,1)x(r)dr = o, (51)
where
KU, 1') = '('r t) + (1' - t). (52)
The convertibility of the problem of stability at least in its
rst order aspect into an examination of linear systems is of in
estimable value since the largest part of existing mathematical
knowledge relates to such systems.
CONCEPTS OF STABILITY
1938) H. \Vold, A Study m the Analysis of Stationary Time Series (Uppsala, Sweden,
Cl. the reference given in footnote 43 See also the brilliant contribution of
Professor Ragnar Frisch to the Cassel volume, Propagation Problems and Impulse
Problems in Dynamic Economics, EconomzcEssays in Honor of Gustav Cassel (London
1933). pp 171205. While ordinary periodogram analySis will not do the Generallzed
Harmonic Analysis" of Norbert Wiener is designed preCisely for problems of this type
See the references to his 1930 Acta Mathematics article in H. T. Dams, The Analym of
Economic Time Series (Bloomington. Indiana Principia Press, 1942), and 3130the refer
case there to the 1935 article of Barrels. and the latter's suggestive concept Of the
"harmonic dial "
F UNDA 11!EN TALS OF D YNAMI CAL THEOR Y 343
above. It is clear, however, from the work of Slutsky 5and others
that the time series generated by such sequences resemble quali
tatively the usually encountered economic time series.
By means of the analysis of the previous chapter it is clear that
(except for terminal adjustments which become negligible in a long
series) the solution of the damped dynamic stochastic system of
the following type
L(Y) = Y(t) +a1Y(t 1) + +anY(t 72)= {z,}, (55)
where z is a random variable serially uncorrelated in time and
drawn from an unchanging universe whose rst two moments
(0, fr;-)exist. takes the following form
Y(t) = A0Z(t) +A1Z(t 1) + +AnZ(t -n) + (56)
As t grows the number of coefcients in this sequence becomes in
nite, but in such a way as to leave their sum and the sum of their
squares nite and equal respectively to (1/2": a,,
le). By the usual
0
central limit theories it is easy to show that (except for terminal
adjustments) the variance of Y(t) is given by
v, = (A02+ As + --+ ARM. (57)
or in the limit by
Ven: k}, (58)
where k is nite. The actual frequency distribution of Y(t) tends
to some limit with increasing t, with zero mean and variance equal
to the last given expressions. This is usually true even if the orig
inal frequency distribution of z is not at all normal (i.e., Gaussian).
The interested reader can satisfy himself that the superposition
of a random variable on to a periodic force leads to a forced motion
Eugen Slutsky, "The Summation of Random Causes as the Source of Cyclic
Processes," Ecanametn'ca, V (1937), 105-146.
The article by Trygve Haavelmo, The Probability Approach in Econometrics,"
Econometrica, vol. XII, Supplement (1944), deals with the problems ol' determining
empirically such stochastl relations. The article by H. B. Mann and A. Wald, On
the Statistical Treatment of Linear Stochastic Difference Equations," Ecanometrica,
XI (1943), 173220,shows that the conventional least squares treatment of autocorrela
tion is (asymptotically) a consistent" method of detemiining the a coefcients. It
would be outside the scope of the present work to go into these problems. A number
of articles in the Annals of Mathematical Statistics for 1942 deal with the sampling
distribution of the autocorrelation coefcient.
344 FOUNDATIONSor ECOAOMICANALISIS
Just hke that described above except that the mean of the asymp
totlcally normal distnbutxon uctuates according to the penodu:
functlon descrlbed 111the prenons sectlon He can also ork out
the lmphcatrons of havmg a stochastic variable 2; whrch IS not
serially 1ndependent
(66)
}; lefuthjz, ; O (8)
It ISnecessary and sulcnentfor a regular relative maximum that
fo : a (t : lt : )
and
);, 2f,,h,h,
1 <0 (9)
for not all h's equal to zero ln addltlon to the vamshmg of all
rst partial denvatlves a regular relatwe mammum requires that
the quadratlc form whose coefcxentsare the second part1al denva
tlves be negatwe denxte Thrs last condltron 15the generahzatlon
to many vanables of the Singlevanable condmon that the second
denvatxve be negative for a regular relatwe Simplemax1mum Its
meamng mll be amplred m a later sectron
The proof of these necessary condxtlons can be easrly demon
strated by the use of a devxce Which enables us to convert our
problem mto that of a single vanable one upon whaeh we can
bnng to bear the results of the prevnous sect1on
Let
x. = xg + u, (z = l, , n)
here the h's take on arbitrary values Clearly for a relatne
maxtmum xxemust hue
f(x10 + Il; xzo + th! , xno +1,13) f(xlov vxno)
MAXIM/l AND QUADRATIC FORMS 361
For given values of the hs, not all equal to zero,f can be regarded
as a function of the single variable t, and the above condition can
be written
f(t) f(0)
From the previous section we know that this requires that
f(0) = 0,
and
f(O) ; 0.
But, by performing the indicated differentiation, we nd
fn) = ; fijohihj % 0.
In order that (10) be satised for arbitrary values of the hs, each
partial derivative must vanish, or
fio=0. (z'=1,--',n)
For a regular relative maximum, we get by this method
fe=0. ('i=1,'',n)
and
6"(361, , xn) : 0
Under what condltlons wnlla pomt (X) m S afford a relatwe
maxzmum to z, and satlsfy the subsndlary cond1tlons? Such a pomt
ISdened as follows
f(xlp , xn) : f(xloo ! x00)!
where
G"(x1, , x,.) = G"(x;, . x,?) = 0, (k = 1, ,m)
and (X) sucxently close to (X )
The way to solve thls problem whnchwould suggest itself most
naturally would be to use our m subsxdxarycondxttons to ehmlnate
m of our variables by expressmg them terms of the remammg
(n m) vanables We assume that the rank of the matnx
x. 6x; 6x.
& 92 66
83: x a ,.
x , or [G] (k=l, ,m)
(!=I, ,")
@ af: 86"
6x1 6x2 ax
ISof order m so that at least one m2determinant does not vanish
MAXIMA AND QUADRATIC FORMS 363
This insures us that we can in fact uniquely solve our subsidiary
conditions for m variables in terms of the remaining (n -m).
Since the numbering of our variables is arbitrary, we may write
this without loss of generality as
xi = f(x+1, xiii-{~21 ' ', xn)- (1: = 1, ' ', m)
_QL 66*
x. x. = :d) GS
and so forth for all possnble combmatlons of (n) varlables taken
(m + 1) at a tlme These wnll yield (n -m) independent condx
tions on the partial denvatwes off and the G s
For concreteness consxder the case of a many variable function
to be mammlzed as of one subs1dnary condmon of restraint Ac
cordmg to the condmons mdlcated above there wxllresult at an
extremum tangency between the locus of restramt and a contour
locus of z (11 l) condmons famlhar m many branches of eco
nomlc theory
In addmon to the rst necessary condmons upon rst partial
denvatxves there are as m the prewous sectlons certam added
secondary condmons on the second partnl denvatues The fol
lowmg must hold
iinm
l 1
n k.
z=o ( .)
M'AXIMA AND QUADRATIC FORMS 365
In summary, it is necessary and sufcient for a regular relative
constrained maximum that
Hi0=01 (1;=1,"',1l)
and
i flijollihj< 0
HM: 1
for
n. k
2 (??]!
- =0 not all hs = 0.
_ Za
_ _ 2by :1: \i4b2y2-4acy2 : [- b :}: @
Vb? 06])!
But this must yield no real value of x or the form would not be
positive denite. Hence, the expression under the radical must be
negative, and we have as an additional necessary condition
ac b2 > 0. (21)
However, our three necessary conditions (19), (20), and (21) are
not all independent. The last of them in conjunction with either
of the rst two implies all three. The sufciency of these condi
tions is implied by the fact that the form is positive for the values
(1, O), zero only for (0, 0), and never changes sign; consequently,
it is positive denite.
These conditions may be given a more convenient form. The
matrix of the form in two variables is obviously
a b
[b cl
If we dene a principal minor of a determinant as a subdeterminant
formed by crossing out the same rows and columns so that the
diagonal elements of the minor are contained in the diagonal ele
ments of the original matrices, then our necessary and sufcient
conditions are implied in the statement: all principal minors must
be positive. These are three in number, but not being all inde
pendent, a complete set may be specied by making successive
principal minors positive starting with the upper left-hand corner
and adding successively corresponding rows and columns. (In this
case there is only one row and one column to add.)
368 FOUNDATIONS OF ECONOMIC ANALISIS
{or the form to be neg'xtwc denite ll. could be shown along
sumzlarlmes that
a < 0
ab
bc > o. (22)
or that the rst prmcxpal mmor ISnegatwc, and the successwe ones
alternate m Sign those of odd order being negative, and those of
even order bemg posmve Thts ISbecause [a ,] nsnegative demte
only 11'[-a ,] ISposnttve demte
Farms m any number of tenable:
The specnalmethods applicable to forms of few variables cannot
be apphed m the general case We must therefore seek more
general methods A non smgular transformation of varrables of
the type
h...b.,p,. (1:1 ,n)
or usmg matrm notation,
h = bp, p = brh,
leaves the quadratxc form posrtwe demte If ongmally posxtne
demte Tlns ns made plausuble from the recognition that the
ongm goes mto the Ongm, and that every pomt m the h space xs
umquely related to every pomt 1n the 13space Consequently, Q
after the transformation takes on only the \alues tal-.cnon before
00!) = (2(1)?) = @(P)
In partzcuhr let the transformation 6 be such as to take Q mto
a sum of squares of the form
Q(bp)= aw + dm; + + (hp,? (23)
lt ISwell known that an mmty of transform'ttnons eust winch ml!
do the
In matrix notation
Q = Il ah
= p (6 ob)? where h = bp
6 must be chosen so that
; 0 0
0 a, 0
(25)
Mzmmumpropane: of Q
Some preliminary observatxons are order before dlscussmg
exphcxtly the second transformatnon Because a quadratlc form
:s homogeneous of the second degree the hs 1f1ts Slgn ls always
posxtxveon the umt hypersphere dened by
hf + ha: + + hu= 1. (26)
1t 18posmve everywhere except at the ongln There IS no loss
m generahty therefore, m consndermg the sxgnof Q subJect to the
s1de condmon (26)
Smce Q 19everywhere coutmuous thh continuous derivatives
1t necessanly attains a mlmrnum on the closed hypersphere For
the form to be posmve ct 1s necessary and sument that the mzmmum
of Q on the mnt hypersphere be posztwe,because then, and only then,
IS1t posmve everywhere except at the ongm
This mnmmum can be calculated expllc1tly Cons1der
X _ '
:
Then the pom: (fu/r , [i./r) :s on the unit hypersphere and
5 1
Q(;) "' ", Q)
ao that then sxgnsagree
MAXIMA AND QUADRATIC FORMS 371
subject to
?!
Zhi21=O.
l
Form the expression
]: .*]:." Il 51
1
' lt can be shown that A. ISthe mnmrnum olQ o l the set of pomt: lying on the umt
hypersphere and sausfymg the oonstramt def ned !
'n')" + n]! + + b,,I:. = 0
If we adJom still another constramt
131+ 59% + + 13.313.= 0
A. mammazesQ Generally x. mnmmazes Q subject to
Ein-=:
(2(1))= z a- ?. ';hfpkhm
" 1|
ll
HIV]: 12 (12 aijhikhim)pkpmw
% we. (33)
2 Z aiihfhs'" = M 2 113115
= Ab
1 l I
(36)
(37}
a1 =
:12 in."
A A
where A is the determinant of the matrix a, and Ar,-is the cofactor of the element in the
ith row and jth column of the determinant A = Iagil- a"1 is termed the inverse of a
beuse the transformation a" following the transformation a is equivalent to the
identity transformation I because of the property
ac = a"a = I.
Only non-singular matrices have inverses.
1Sylvester's Law of Latency states a more general relationship. The latent roots
of f(a) are equal to f(x) where the A'sare the latent roots of a. This followsfrom the
376 FOUNDATIONS or ECONOMIC ANALYSIS
V. QUADRATICFoams Drmun. UNDER CONSTRAIN'rs
Let us now derive necessary and sufcient conditions that
quadratic form
does not require as secondary conditions that the matrix of the Hessran
[Wii]
be a negative denite form In fact, the demteness of this matrix ls no! invaria
under a general monotonic transformation of the utnhty mdex of the form
F = F0?)
F; = Fm
Fu : F.9d + F".,,
ms the correct conditions later to be developed are Invamnt.
M'AXIJHA AND QUADRATIC FORMS 377
. As in the previous section, there is no loss of generality in con
Sldenng the behavmr of Q on the unit hypersphere dened by
>?sz 1 = 0. (42)
This additional side condition is adjoined to the m equations of
(41), and Q is to be minimized. Form the expression
ngh; = O, (a = 1, - - -, m) (45)
hf: ' 1 = 0.
1
R : 031
g ann
g gal
0 gum
0 __ [ga _g]
0 (49)
gl 3n" 0 0 J
The matrix has been part1t1oned mto four submatrlces, one 151:
by n, another m by n, stlll another n by m, and the last m by m
The matnx ISsymmetncal, 1e ,
R = R' (50)
The form xs negarwc demk 1f and only nf all are negauve If the root: (1Her
sngu It IS mdemte If one or more are zen: but the rest are of the same sagn xt :s
.MAXIJIIA AND QUADRATIC FORMS 379
The inverse or reciprocal matrix is dened by
RR1 = I = identity matrix.
It can be written
[Mud3H
g 0 ' q ' (51)
[ad+gla+g9]=[f__0_]_I
gd+0 lg+0 0 I (52)
A SmGLE EQUATION
2 A dl'erence equation can be written as an 1mphc1trelatlon
of the followmg form
fD'U + ") :'(! + n - I) .:) t] ==0 (1)
Tins W111be said to be of the nth order because nt mvo1ves the
variable y at n dlerent (equally spaced ') penods of time It may
be noted that the fact that the y yalues are dened for t values
differing by an mteger does not really Imply any restnctzon for
1fthe equally spaced values dlffer by any quantxty, we may redene
vin urne vana'oxe so as to make fms 'erence equal to nmty
Although the values of t dlffer by an Integer, lt 15not necessary
that I take on only Integral values However, for purposes of eco
' The econormst may consult wnth prot G Boole Treame on the Calculusof Funk
D!!! enccs (London 1880) L M Milne Thomson The Calculus of mete Dterences
(London 1933) P M Batchelder An Introdudmn :o Lmear Dzcrence Equation!
(Cambndge Massachusetts [927) N E Norlund Vorlewngen aber Dzcrenenrechmmt'
(Berlm 1924) for a more advanced treatment and blbhography
-DA
DIFFERENCE EQUATIONS 381
nomics it is usually only necessary to consider the values of y for
integral values of time. By breaking up our time into small enough
units we can keep the error involved in this procedure down to any
requisite level. As we shall see, this greatly simplies the mathe
matical theory and makes it resemble very closely the theory of
differential equations.
Equation (1) is an identity in t. The reader will see from this
that it may be written in very many ways, e.g., as a single relation
not between [:y(t + n), - -, y(t)], but as a single relation between
[y(t), - - -, y(t n)], etc. However, because equation (1) contains
as its last factor the explicit term t, rewriting our difference equa
tion in the manner suggested will involve a slight modication of
the form of the f function. Nonhistoricalsystems, which are of
the greatest interest in dynamic economics, do not involve time
explicitly and therefore omit the last I!in equation (1).
3. Difference equations receive their name because they involve
the value of a variable at different points of time. As written in
equation (1), however, there is no direct use made of the notion of
nite differences My, dened in books "'on interpolation as follows:
AW) = W + 1) - y(t),
NW) = AW + 1) - AW) = M! + 2) - N + 1)}
{W + 1) W)!
= W + 2) - 2W + 1) + W),
(2)
= Z ni')!z.!(])
(n- y(t +12 i)
Ay(t) = W)
It will be intuitively clear that a knowledge of n adjacent values
of y will permit us to calculate the value of y and of its rst (n 1)
differences. (See equation 6 below.) Therefore, by substitution
into equation (1) we can always write it in the following form:
FEMS), NIDU),' ' '! 43'): 30), t] = O, (3)
2Sec E. T. Whittaker and G. Robinson, The Calculus of Observations (3rd ed.,
London: Blackie, 1940), chaps. iiv.
382 FOUADATIONS OF ECONOVIC ANALYSIS
where F 15related tof as follows
PIA'y. . Ay.y. t] = ID + my
+Lzquy+ +A v tv+AyvyotJ (4)
Not only can every dn'erenceequation hke (I) be wntten as (3),
but also conversely, because of the relations gwen m (2), every
equation of form (3) can be ntten as form (1) Thus the mo
forms are completely equnvalent, and rye can choose bet een them
on the basxsof com emence Actually, the form of equatxon (1) xs
most frequently encountered m the hterature, although later It Will
be shown honr either may be handled
4 Thetr relattonshlp to each other may be further revealed 11
e mmme bney the notlon of an operator Thus, A may be
thought of as a dtfference operator dened by the rst equatnon
of (2) In a Slmxlarway he may dene the dlerentlal operator
D(= d/dt) by the relatxon
Dy(t) ==y'(t). DW) = Dy'G) = :v"(t). etc
In the last century the English mathematman Boole, followmgup
the earlier fruitful suggestion of Lelbmtz, employed the shnltmg
operator E dened as follows
EN) = W + l). EM!) = y(! + 2). etc
The successof the operator method stems from the recogmtxonthat
these may be treated as 2)"they ere algebra1c quantmes Thus,
leavmg out the 3: function completely. we may wnte out the
operator Identity
E = 1+ A (5)
Treatmg these as algebranml etpresszons subject to the usual b1
nomlal theorem we deme Immediately the relations
En=(1+A)*=1+nA+'1-(2_Qm+
n A
(??')
=(E1)=E"-nE""1+-E"2+ + '(6)
l
+(_ I) 09,15
(11 +
ate the agreement of the second of these 1th equatlon (2). and
the use of the rst of these m connectxonmth equatton (4)
DIFFERENCE EQUATIONS 383
Exercises:
1. The algebraic deviation of the price of hogs from the equilibrium value
is in any year %the deviation of the previous year, but of opposite sign. Set
up a difference equation for the price of hogs. Intuitively is the process stable
or unstable? (Hint: Use as your variable the deviation of price from the equi
librium value.) \Vrite out the rst few terms of the sequence.
2. Form a difference table of the numbers 2, 4, 8, 16, 32. How many
difference columns can you ll in with the given numbers? What is the general
rule relating the number of columns of the difference table which can be derived
from a given set of numbers? Note the triangular form,
3. Write the difference equation of the rst exercise in the form of equation
(3) rather than (1).
4. Write the difference equation y(t + 2) a(1 + B)y(t + 1) + ay(t) = 0
in terms of the E notation and also in terms of the A notation.
5. What is the meaning of E?
6. The receding difference operator V is dened by Vy(t) = y(t) y(t 1).
Give its relationship to E, corresponding to the relations of equation (S).
7. Write out the usual Taylors expansion for y(t + 1) in terms of y(t)
using the .D notation. Verify the relation 2 = E. (Hint: Use the series
expansion e = 1 + x/ 1! + x2/2! + ' ' '.) What would seem to be meant by
sine D? by 1/(1 A)?
8. What is the rst difference of ?? the second and third differences? What
is the third difference of any second degree polynomial? What is the nth
difference of any nth degree polynomial? the (n + 1)th difference of an nth
degree polynomial?
9. Write out the difference equation satised by the numbers in exercise 2.
10. Let us dene ascending factorial polynomials as follows: :d") = x(x 1)
- - -(x n + 1). Verify that Ax") = nx"'. Of what does this remind you?
SYSTEMS OF EQUATIONS
5. Up until now we have been concerned only with a single
variable. In most applications to an economic system many vari
ables are involved. Hence, we must introduce the notion of a
system of difference equations in m variables (yl, - - -, y). These
consist of a set of difference equations in the variables in question.
Just as a static system can only hope to be determinate if it consists
of as many equations as unknowns, a dynamic system must contain
as many independent, consistent equations as there are unknowns.
Thus, our system takes the form
Ey1(t+ n).y2(t + n). ---,ym(t + n);
3'1). 3'2): ' " 391.0);1 = O,
(7)
(i=l,'-.m) (Il)
Byte1.: = ni
BY,; = J.
Our initial conditions are written simply as
[YMOD = [Yd]. '1):|- (12)
By way of illustration let us consider a system consisting of
one second order equation in one variable.
N + 2) + a:y(t + 1) + my) = sin :; y(0) = bo, y(I) = b1. (13)
The order of this system is clearly two. Dening new variables
as follows
Y!) = ya)!
YA!) = 53(1). (14)
we can write the above system in the normal form
BY,) = _ (Y,) _ Onyx)+ sin !.
Elli) = Y!).
Yl(0) : bo.
(15)
Y:() = b].
The reader should work out several examples, such as that of
exercise 4, section 4 above.
DIFFERENCE EQUATIONS 387
It is to be noted that equations (7) can by a similar denition
of variables be written as S rst order implicit equations. Indeed,
an indeterminate system, involving more unknowns than equa
tions, can always be converted into an (equally indeterminate)
rst order system.
9. We have seen that a single high order equation can be trans
formed into a number of rst order equations. Is the reverse
possible? Can a rst order system (and, therefore, any system)
be transformed into a single equation, whose maximum order is
equal to the maximum order of the normal system? Usually the
answer is yes, although the reduction is easier to indicate than
actually to perform, and in singular cases, where certain conditions
on the partial derivatives of the fs fail, the reduction may be im
possible. Also, a well-determined explicit system may after reduc
tion be expressible only as an ambiguous implicit high order relation
in one variable. Thus, the well-determined system
By: = 3'12; Ey1 = yz + yz (15)
can be reduced to the well-determined second order equation in y1
Eyx - Eyn = 3:12, (17)
but in yz we get only the ambiguous implicit equation
(Ezyg Eyg y} 2 == 4y22Eyz. (18)
Let us suppose that we have a normal system of order S,
ya + 1) : JiEyl(t)v ' ! y-(t): t]! (i = 17 ' ' '1 S) (19)
or, in abbreviated notation,
EY = J (I, t).
It will necessarily involve ZS variables, namely the y's at two.in
stants of time. Given S equations in ZSvariables, it is notpossrble
to eliminate all but one of the ys and be left With a difference
equation involving only that y. If we write down the identities
E2Y = J(EY,t + 1)
EY = J(E2Y, t + 2) (20)
' The reader may consult F R Moulton, Dijermtial Equation: (New York, 1930).
pp 69.for a more complete discussiion of this problem taken Wlth specul reference to
differential equations For linear equations this is worked out in some detail in P A
Samuelson, A Method of Determining Explicnly the CoeffiCientsof the Characteristic
Equation," The Annals of Mathematical Statistics, XIII (1942), 424-429
DIFFERENCE EQUATIONS 389
5. A body falling in the earth's gravitational eld satises the equation,
(Fac/dt2= g. Convert into normal form.
6. Write out a system of di'erence equations in A form. Dene the con
cepts of order, normal form, etc.
11. If only numerical answers are desired, one may simply use
the difference equation at each instant of time to calculate suc
cessively the new values of the variables, and so forth for as long a
time interval as is desired. However, this may be tedious in prac
tice, and inevitably there must be rounding off at each stage so
that the accuracy of the solution deteriorates with its length. (For
a stable system if the errors of rounding off remain within a certain
limit, the errors of the solution will remain bounded.)
The method of actual numerical examples has had considerable
use in economics, and is valuable from the standpoint of providing
illustrations. Still one can never be sure that another arithmetical
example might not contradict the principles inferred from this one,
so that the method is of limited usefulness even for the purpose of
providing qualitative information concerning solutions. After all,
life is not long enough for us to complete even one sequence, nor
to try all possible initial conditions, nor to vary the form of the
structural relations.
Unfortunately, if we simply consider difference equations in all
generality, there is no simple analytical way to write out explicit
solutions. The same is true of differential equations. The be
ginning student may be taught a variety of ad hoctricks which may
work on particular equations, but the advanced student must rest
content with his existence theorem which tells him that there exists
a unique solution; often the proof of the existence theorem tells
him how he may (painfully) approximate to that solution.
In both of these lines much more complete results may be
achieved if we restrict ourselves to linear systems. That this may
be a serious limitation, especially for business cycle purposes, is
shown in chapter xi. However, in chapter x it is shown how the
assumption of linearity is not seriously limiting if we are primarily
interested in stability, especially in the small.
In much of what follows, therefore, we shall conne our atten
tion to systems which have two properties. First, they are to be
linear. There is the added advantage that most mathematical
E. Lundberg, Studies in the Theory of Economic Expansion (London, 1937).
390 FUUNUA'IIUIV.)
Ur [nu/"Lump annular.)
exposmons are concerned wrth tl ese Second they are not to an
volve tune exphcrtly Most economic systems of mterest are
causal rather than hlstonml m mture and satnsfy this require
ment thereby szmphfylng our task and enhancnng mathematlcal
tracttbrhty
In fact 1!we msust that a system be tnneless and at the same
t1mehne1r we are left wnth the exceedingly Simple case of constant
coeiczentr whose theory :s perfectly complete and elementary
We shall see that euctly the same thmg lStrue of differential equa
trons so that one who has mastered either of these branches has
httle drmculty Withthe other From a more fundamental pomt of
we each of these ISSimply the exemphcatlon of a more general
theory of llnear operattonal equations
12 There 15 however one nan lmear case n here a more or less
complete solutron can always be md'tted That Is the case of the
rst order equatlon of the form
y+1 =f(3':) (21)
13. The case of the rst order dierential equation which does
not involve time explicitly,
dy/dt = f(y). (26)
is one which admits of an immediate solution by quadrature, as
indicated in chapter x. Its richness of solution is limited, every
motion being monotonic. Hence, if bounded, the motion must
approach an equilibrium point. There may, of course, be a number
of equilibrium points alternately stable and unstable.
Exercises:
1. Let f(x) = x. What is the general expression forfdx)?
2. Show that every equilibrium value of f(x) is an equilibrium value of
f2(x). Is the converse true? What about the stability of the common equi
librium points, will they agree?
LINEAR SYSTEMS
14. A linear system is one which involves the dependent vari
ables in expressions of the rst degree, i.e., without higher powers,
cross-products, nonelementary functions, etc. For the purpose of
this section it is not necessary that time be excludedor that it enter
the problem linearly. We may write our linear system in the form
LDUH = f(t), (27)
392 ['UUJVUIIJJUJJ.)
ur Luvuuuzzu annular.)
where L stands for any general operation upon a whole system of
vanables and f represents a known functlon or set of functions
For the dlscussron of thlS section It IS not even necessary that L
represent dxerence equatlons rather than d1fferent1alequatlons
Rather ntmay represent an lmphclt functional of y over the whole
I Interval and In the notation of chapter m page 261 It could be
wntten as
LMD] ==PDG}: t]
The fundamental lmear prOperty which L must satrsfy [S as
follows
LU: + yz) = LCM)+ Lfyz) (28)
from WhIChfollows Immedrately the general relatlon
L(a1y1 + 02y: + + anyn)
= 01L(y|) + 02L(y2) + + anL(yn) (29)
From this general demt1on of a lmear system there follow 1m
medlately two Important general theorems
Superposmon theorem The salutwn of (27) when f 15made up
of the sum of two dz'erentfs :: equal to the sum of the solutzons op
propnate to each Mathematically thls may be wntten m sllghtly
more general form
Msn) = I lmphes HEM!) * ZaJ (30)
The proof follows immedrately from (29)
Thns means that the solution of a system wrth many forces
actmg on 1t IS equal to the sum of the separate effects of each
Even where tlns ISnot stnctly true the economlc sxtuatlon may
approxnnate to tlns condltron
For many lmear systems It may be posable to gwe the solution
for a fundamental set offs mto whrch any f can be decomposed
If thls ISthe case our problem 15Immedrately solvable
Our second theorem states The dz'erencebetween any two salu
tzons of (27) must satzsfy the reduced (or homogeneous) equatzon
LU) = 0 (31)
T1115follows Immediately If we substxtute each of the solutlons m
77) and then subtract one expressron from the other The right
DIFFERENCE EQUATIONS 393
hand function will cancel out, and according to (29) we may regroup
the terms on the left-hand side to achieve the desired result.
A most important corollary of this theorem is derived simply
by rearranging terms. The most general solution (i.e., any solution)
of (27) may be written as the sum of some one particular solution and
the general solution of the reduced equation (37).
This permits us to concentrate upon the more easily manageable
reduced equation in order to determine the uniqueness, multi
plicity, boundary conditions, etc., of the solutions. It is often
possible by ingenuity, good luck, or intuition to arrive at a par
ticular solution of the unreduced case, and the present corollary
tells us how to transform it into the general solution, if only we can
nd the general solution of the reduced case.
Exercises:
1. Except that there are no boundary conditions, show that the above
remarks apply to a single linear algebraic equation of the form, ax = b. Show
that when the reduced equation, ax = 0, has but a single solution, the full
equation has but a single solution. Can the converse be stated? Can you
state that the uniqueness of either equation guarantees the uniqueness of the
other? (Hint: Distinguish three possible cases, neither a nor b equal to zero.
both equal to zero, only one equal to zero.)
2. Consider two (or more) linear algebraic equations. Show that unique
ness of either the reduced or unreduced equations implies uniqueness of the
other. Show, however, that the reduced equation can have many solutions,
when the unreduced has none. Enumerate all the possibilities, and with the
aid of a textbook on algebra show how one can recognize each case. (Hint:
The terminology of matrices and determinants will be found useful.)
3. Show that the following integral, and mixed dierence-dierential
equations are respectively linear:
M + f; ku mane =f(t),
y'a) + ay(t 1) =f(t)
Do they involve time explicitly?
330) = 0, t g 2,
is a solution of (38) for appropriate initial conditions, and yet it
cannot be written as a linear combination of the two linearly
independent solutions.7
7l t may be objected that 330)does not satisfy the difference equation (33) for all
values of t, particularly for ! equal to zero. However, a solution is to be thought of as
satisfying a difference equation if for given arbitrary initial conditions it is in accord with
the prescribed law. To help clr up the present paradox, it should be pointed out that
solutions which are linearly independent on one interval of t may be linearly dependent
on another interval. It is also to be pointed out that the example cited is one in which
the solution of the difference equation is unique to the right of the initial conditions but
not to the left. It is also one in which the irreversibility in time disappears if the differ
ence equation is replaced by an equivalent" one of lower order.
396 FOUNDATIONS 0r ECOAOIHCANALISIS
The absence of a general solution does not me m that we cannot
nd the solutlon for gwen mmal condmons Our existence
theorem guarantees us tlnt ne can nd such '1solution Actually
such a solution can be \\ ntten ISa lmear combmatnon of basnc
solutxons However the aboxe evunple shows us that the linearly
Independent basic solutions may haxe to be dt'erent sets for every
mltnal pomt there need not cust a set of basxc solutions which wrlI
serve umformly for all mltral times
LINEAR OPERATOREQUArrows wmr Covsmvr
Corrrlcrrms
18 I turn now to the case of drfference or ddl'ercntnl cquatrons
With constant coefcrents Since xt ml! be shoxm later that the
solution for equations of any order can be written terms of
solutions of the rst order C'lse I shall consrder thus Simplest of all
cases In the rst order case our drfference equation becomes
(E mm = K!)
yUo) = b (4)
and the correspondmg differential equatlon becomes
(D - 9M!) = f(t)
m =b 42)
Each of these may be consulcred as specral exemphcanons of
a general rst order equatron m a Simple opentor 11
(h " 9)y(f) = f (f)
?(10) = b (43)
For h to be an admzsszbleoperator nt nsonly necessary that there
be an emstence theorem guaranteemg that the above equation has
a umque solutnon for the arbltranly specred Jamal condition
How thrs exrstence theorem rs estabhshed need not concern us at
this pomt we hu e seen how it can be done for differenceequations
and the reader may refer to any orL on dl'erentnl equations for
the corresponding proof In addntron to E and D examples of h
are A 0 E + 5 etc On the other hand such an operator as D
would not be an admrssrble h because the system correspondmg
to (43) would not have a umque solution Thus :s because mo
lmtIaI conditions are necessary rather than one to satlsfy a second
order dxfferentlal equation
DIFFERENCE EQUATIONS 397
.We may . suppose the solution of the above system to be
written as a linear functional of f, with q and b as parameters, or as
F(t, g, b; f)where F has the fundamental linear property
FU, g, Clbl + ' ' + cabs;lel + ' ' ' + csfs)h
= GIF! q, bl;fl)h + ' ' ' + CSF(tyq: b3;f8)h' (44)
From this fact it is easy to verify that h is a linear operator
with the properties
[t(Clyl + 62372)= Clhyl + Call,
Iz2 = lz(hy),
' (45)
h"y = hh"y,
h"! = ly,
110 = 0
P(/1)Q(h)y = Q(h)P(/t)y.
where P and Q are polynomials.
19. So far we have not dened It raised to a negative power.
It would be natural to dene
y = hf(t)
as the solution of
by = K!)
However, it is clear from our existence theorem that this cannot
have an unambiguous meaning unless we specify some initial con
dition. This corresponds to the familiar indeterminacy of the in
denite integral of a function, Dlf. We need a constant of inte
gration or a denite lower limit of integration before there can be
a determinate answer. We shall not have occasion to use negative
powers of It even though readers familiar with the Heaviside opera
tional methods will realize the mnemonic advantages to be derived
from such a procedure. These advantages are counterbalanced by
the aura of mystery which surrounds some of the usual expositions
of this method. In general, the methods here outlined will be of
classical type, albeit in somewhat amplied form. It is the
present writer's conviction, which need not be sustained here, that
a careful examination of the Heaviside-Cauchy operational calcu
398 FOUNDATIONS OF ECONOMIC ANALYSIS
lus, of Bron-m1chWagner contour Integrals, of Laplace transforms,
etc , Wlllshow that where these dlfer from the classxcal methods
the advantages are m favor of the latter These remarks are
tended to apply to ordmary dt'erenttal equattons of mte order
Wlthconstant coccxentsand do not refer to the use of the Laplace
transform 1n connectxon w1th partlal dt'erentlal equatlons, nor m
connection w1th Integral equations of faltung type, etc
20 chg rst concerned wnththe reduced equatlon, wheref(t)
vamshes we concentrate upon solutions of the form Fa, q b, 0)
Because of the hnear prOpertIes of h gwen m (45), this can bc
ntten m the special form bF(t, 9 1,0). Hence, pecuhar :mpor
tance attaches to the solutnon of the reduced rst order equation
Wlth mutual condntlon equal to umty Let us dctcrmmc, there
fore, the exact form of the F function for h equal to E and for 1;
equal to D
Clearly t(
Ey = ay. We) = I.
then
We + 1) = Wo) = q.
(46)
D ""' e9"*')f(a)da,
J; (52)
ott
E 2 9f(z)
etc
h"y(t)]-r. = b,
has a solutron. and the solutlon nsumquc
To prove the rst part, I shall specxfyone solutlon, leavmg lts
umqueness to later drscussxon FllSt,let us note the algebratc fact
that every nth degree polynomml has exactly n roots If we count
each mth Its approprmte multzpllcnty If q. denotes the 2th root,
we may nte our polynomlal m the form
Pn(x) = (x _ 1)(x _ g!) (x "' a) (57)
Let us dene 1:new polynomials as follows
, (58)
Pf(x) = P?+\(x)/(x _ gran"): ? au rxrt'
(h g,.)ZnU) = Z,(t).
By' our hypothesis on a rst order system, each of these is deter
minable once we specify initial values [Z1(to), Z200), - - -, Z,.(to)].
In a moment these will be prescribed.
By virtue of the denitions of Zk(t) in (59), it will be clear that
f (t) = ZoU)
= (h _ 91)Z1(i)
= (h ' 91)(llgaza) (60)
= (h gl) (h -qn)Zn(t).
Thus, Z,.(t) is a solution of our equation, and it remains only to
make certain that we pick the Z{(to)properly so as to yield the right
initial conditions.
It is easily veried that
Znj(t) = P,-(h)Z,.(t). (61)
Evaluating this expression at the initial time we have
(j=0,1,2,-,n1) (62)
These are sufcient equations to determine the right-hand terms
for a given set of bs.
To recapitulate our construction: nd the roots of the poly
nomial P ; by synthetic division or otherwise dene the n new
polynomials, Pi; evaluate Z,-(to)in terms of the bs; by repeated use
of the methods of the last section, equation (54), solve in turn for
Zi(t). 220), - ~,Zn(t), using the initial conditions just determined;
the resulting Z,.(t) is the required solution. This Proves the rst
part of our theorem.
404. FOUNDATIONS OF ECONOMIC ANALYSIS
To prove uniqueness e shall assume that there are two dls
tmct solutions mdshow that this leads to a contradictlon Clearly
the dl'erence between any tuo solutions of (56) ml] satisfy the
reduced equation wrth :nrtnal conditions all equal to zero We
have only to show that the reduced equatron wrth zero rmtral con
drtrons cannot have a solution whrch usnot :dcntlcally zero For
the rst order case our existence theorem msures that there can
be no other solutron than zero We shall proceed by mductlon to
show that 1f'1reduced equatlon of order (n - 1) has a umque zero
solution [or zero rmtral condltxons then so must an nth order
equatlon
Let us suppose that the (n -1)th order equation .has a umquc
vamshmg solutron and the nth order has a non vamshmg solutlon
Y(t) Then
{PUYl = (h _ &)le uUYl= (h _ quz = 0 (63)
where
hY(tg)=0 (r=1 nl)
YU) # 0
and Z rs dened by thrs equatron Our hypothesrs on the (n 1)th
order equatlon s umqueness tells us that Z must be nonvamshmg
smce otherwrse our nonvamshmg Y would satxsfy the (n 1)th
order equatnon On the other hand the zero values of the n mrtral
condntrons of Y make Z(tu) be zero Wrth 'un 1n1t1alvalue of zero
and satlsl'ymg a rst order equation Z must vamsh Identically
Thus we hate 3.contradnctnon and our conclusron concermng the
uniqueness of our solution rs proved
It ls to be noted that the case of repeated roots m the character
lStICpolynomial P Introduces no dlfcultxes 1n the problem Also
It 15to be observed that the method of nding a solutlon outhned
above ISnot necessanly a good one to follow m practical numerical
computatron However smce xtdepends only upon repeated solu
tron of rst order equations rt can be readily mechamzed or reduced
to a routme ll some surtable method of solvmg rst order equations
18at hand (e g dterentral analyzer punch card equrpment etc.)
Etam ses
1 Wnte out the soluuon for a reduced second order (I ercntral equat on
With constant coefc ents wh ch has y(0) = l y (0) 0 vh ch has y(0) = 0
1'(0) = l
DIFFERENCE EQUATI ONS 405
2. Write out the solution of the unreduced equation with vanishing initial
conditions.
3. Repeat (1) and (2) for any operator h, using the F function.
lz"F = q"F,
and
P(h)F = P(q)F. (67)
Because of the linearity of our system, we have
2,, Pht,2Pht,1)
P(h)[y(t.4q)2-:l(19)]= ()y( gaiq} ):v( 9 _ (68)
406 I'OUNDATIONS OF ECONOMIC ANALYSIS
If y is a continuously differentiable function of the parameter q,
we find in the limit
a am,1
CrOFU,gr) + cf] FU; gr) + ' ' ' + Cram15E: FU, qr)
r mkl ai
= i=1
Z 1=o
Cl:.i',-
g 170,91). (73)
where (gl, gg, - - -, q,) are the distinct roots of P with respective
multiplicities (ml, m2, - - -, m,). This linear combination is a solu
tion of the reduced equation. It remains to determine the n co
efcients (Cm, - - -, C) so as to satisfy the n arbitrary initial
conditions in (56). Thus
. f inkl a). _
ht:l=l 1:0
z ijqu(t:Qk)=biy
8 (1:0,1,"',n_1) (74)
or because of (66) and (69)
r ml J'
ZMWw(QkWW(W
t:! j=0
where
dj . . il __. . .
. = .* (j) = -.-. ;' J, ?:
[dg] g :lq-qg (9L) (1_ J)! gl ( ])
=a <p an
408 FOUADATIONS OF ECOAOIHC ANALYSIS
When all of the roots are (hstmct the determmant of tlns
system takes the famxhar Vandermondc form
1 l l
Q: : 9n
m q2 9
= II ((P ... Q:) (78)
:>;
= (79
91 (qt) (W)" " 9: (q: JW 4} (M"
FU! gl! QZ, ' ' '2 gm Oif)h FU! gm 0; Zn-I)h = Zn.
The last is the required solution.
In this section an alternative procedure is described whereby
our Specialsolution with vanishing initial conditions is built up out
of the rst order solution, F(t, g, 0;f);.. First, let us recall the
algebraic fact that any rational expression of the form
1 = 1
($112, = n) (81)
P(x) (x _ Q:)m
!=]
1
A" : Wim-mr:d'" ELLE]
P(x) ==a. (83)
P. A. Samuelson, "Efcient Computation of the latent Vectors of a Matrix,"
Proceedings of the National Academy of Sciences, XXIX (1943). 393-397
DIFFERENCE EQUATIONS 4II
problem. After substitution into the nth order equation we have
P<h>y 1=l
(:1 z A..-[5503]
' QQ (h gain-(z. g. on.
_ r m; P(h) 90
* EaiAjl(hf'qiylf ( )
UU)
by virtuc of the identity (84).
Note that our discussion of negative operators was purely
heuristic and not germane to our proof. The expressionin brackets
does not involve negative operators although notationally it appears
to do so.
This completes our second construction of the unreduced equa
tion with vanishing initial conditions.
27. The elementary mathematical literature on differential
equations abounds with trick methods for arriving at particular
solutions of the unreduced equation for special f's (e.g., poly
nomials, exponentials, trigonometric functions, etc.). At this
point it is only necessary to discuss two such methods.
The rst is a formal series expansion, of even greater importance
in the eld of difference equations than differential equations where
convergence is not always certain. Let us write the formal
expression
1 1 _ . 2 ..
P_(x)=ao+a1x+ +anxno'l'lx'l'zx + '
B = _ (a16,_.1 + (12632
ao+ ' + (1.30) _
412 FOUNDATIONS OF LCOIVOMICANA! }SIS
In pnctxcc there are & mncty of we. of determmmg the 5 se
quence from the 5 (algebraxc dlvzsmn, tltf'ferenee equations
Taylor's exp1ns1onof I/P(x), dctcmumnts, etc)
Forma). a mrtlcuhr solution of
P(h)y = !
ISgncn b) the mmte series
l " ,
= F05! - EONI! (93)
u.here l/P nsno dened differently from (86) and (67) Whether
there ISany mcmmg to the right h'md mic thtt IS hether the
sencs 15dened 1nd comm-gent or not depends upon the h, the I.
and the a's Ilms
(mo + I)? -"-f (94)
has 1 nite com crgcnt solutlon for f eqml to t of the form
) = % ("' a.)'Dc
which com crgcs far INN lcqs than one, but teems to ducrge
othermse Hoxxexer.if we use the "fornnl" adentltv
P(E")y
y(0) =
_ 6110?)
r + 0tny - 1) + ' - ' + 0100 n) = 0,
3'( 1) = .
y(- n + 1) = 0.
2. If P(:c) has n simple roots, gr, and R(x) = ao"x"P(x), show that
" 1
: ___ . (FIn
B. 21;Rm <q.) .
3. If a. equals a/t!, what is B,? (Hint: Consider series expansion of e
and 1/e.)
4. Write B, in determinantal form.
5. Write out the rst three Bs in terms of ds.
c P"(q) (105)
y = f mdm. (108)
In cases where P(q) vanishes, this may be an improper integral
whose limit nevertheless often provides a satisfactory solution.
The mathematician will recognize that in the eld of differential
equations, all the fs that are likely to arise in practice can be
written in the form (107), by setting q equal to ia, and F equal to
ef, and using Fouriers integral theorem.
29. The artice of the Laplace transform will now be general
ized to a wide class of simple operators other than D. The Laplace
transform is dened as
(113)
(q.y)= Eg+ Clg%;)
= P(st).
where R 15a polynomml 1m olvmg the mutual condltxons, 6;, and
ISless than n 1n degree Because I IS known, the Laplace trans
form of our unknown (unctton ts deterrmned Under general
mathematlcal conditions we can prove that our solution y 15essen
tlally" unique once tts Laplace transform is determined One
method of mversxon nsgn cn by contour Integration
1 c+uo
ya) = 53L Y(q)ed9. (114)
where the contour nsto the right of all poles of ?
ln any case, once 7(9) 15determmed, the determmatxon of y(t)
lSa detaxl which need not concern us Extenswe tables are aval!
able to and m mvertmg Laplace transforms 1nvolv1nga wide class
off's
To extend the theory we need only seek a fzmctzonalappropriate
to any admxssnbleh. thh the fundamental property
(q. by) = q(q.;v)h -N) (115)
[he followmg expressron formally ans ers our problem
__ F0 ?! 0k
(q. f),. _ 11m
-"-m.
M 0);. (116)
where 9 and f are such that
f(t)
{LTFU, q, 1.0)h =0 (117)
The proof that tlns 15our generalized Laplace transform ls gwen
us nJlmws
FU qu hf)
(q.hh= [1mm (118)
By denition
(h _Q)F(t!9109f)h 2f: F(lo,g.0,f)h "'-:O
(h q)F(t.q.1.0);. = o, F(lo,q,1,0),. = 1 (119)
(h - q)F(t.q. 0.110;. = hf. F(to. q.0.hf):. ==0.
DIFFERENCE EQUATIONS 417
and each of these systems has unique solutions. Now it can be
veried that
or formally
P1131003.
+ Pm2(h)y2+ + P(hb'm : fn)!
DIFFERENCE EQUATIONS 419
where P1300 are polynomials in h of the nth degree or less; or in
matrix notation as
ooh)! + alhHy + - - anlhy+ any = f @)
and
Py =f,
where (ao, - - -, a,.) are 1122matrices, y and f column matrices of 17:
elements. Of course, many of the elements in the (1smay consist
of zeros.
The maximum order of our system will be nm; but its actual
order may well fall short of this. By the actual order of the system
we mean the maximum number of initial conditions which can be
arbitrarily imposed on the system.12
As an exercise in matrix notation the reader may work out the
demonstration that by redenition of variables 13the system can
be written as a rst order one,
thY+A1Y = F, (127)
where
'I O O Ol
0 I 0 O
A0 = r
0 0 0 I
L0 0 0 ao
r0 I 0 0'
0 0 0 0 0
A1 = r F:
0 0 O I f
ran (Zn1 02 al
" It can be shown that the order of the system, $, is equal to the degree ih q of
the determinant
A(q) = laoq"+ a;q"'l + ' ' ' + anl
See R. A. Frazer, W. J. Dunn, and A. R. Collar, Elementary Matrices and Same
Applications to Dynamics and Derential Equatiom (Cambridge, 1938) for references.
See par. 5.4.
Frazer, Duncan. and Collar, Elementary Matrices. par. 55
420 FOUNDATIONS OF ECONOMIC ANALlSIS
Although our system has been converted to a rst order syster :,
1t15not yet m normal" form thh h Yexpressed exphCItly terms
of Y and functions of tlme
hY==AY+f (128)
If A0 nsnon smgular, that IS.1fxts determmant does not vamsh,
we can premultnply both srdes of (127) by A0 and Immedlately
convert to normal form
hY = (AO-IAM + (A.,IF) (129)
Srmxlarly m (126) 1fao xsnon smgular, our operators of hrghest
order can be solved for m terms of those of lower order by the act
of premultlphcatxon w1th do"
If au 15smgular our system 15not well specred One of three
Sltuatlons must then prevanl (1) there exxstpurely statzcal, algebrmc
relatzansths between some of the vanables By means of these
relatlons one or more vanables can be ehmmated, and the leading
matrix of the resultmg smaller system w111then be non smgular
(2) the relatlonshlps are not all mdependent so that a unique solu
tron does not emst Arbxtrary functlons of trme w111enter Into
the solutron (Example A system mth two Identlcal equatlons,
permlttmg one varlable to be chosen as an arbltrary functlon of
t1me) (3) the system 15mcons1stent
31 It 1s up to the economxst to present a well spec1ed dy
namlcal system In what follows we shall assume that the task
has been adequately performed so that whenever desu'able our
system can be wntten m the normal rst order form of order :
hy = Ay +f. (130)
w1th 1mt1al condltlons
3%) = b
As m prev1ous SeCtIODSwe shall nd our solutlon 1n two steps
rst. the solutlon of the reduced or homogeneous system
hy = Ay. (131)
3'00) = b
and then the solutron of the unreduced system
y(to) = 0
DIFFERENCE EQUATIONS 421
The sum of these two solutions will be the solution to our general
system (130).
32. Earlier we have seen that the solutions of the simple rst
order equation in one variable provided the building blocks for
the solutions of more complicated cases. We try, therefore, a
solution of the form
yl = CIFU:921:0: : CIF, g),
(133)
An An: A _ q
= n'-og.+ fig.l + + 'alq+ Tl = 0 (136)
For a q chosen at random the determmant ml] not vanish, and
ne shall heue no non tnvnal solution But q 15at our dlsposal, we
select xt to make the determmant \amsh The determmant ls a
polynomxal of the sth degree m q, and 15called the charactcnsc
equahon of the matnx (or secular equation) The roots of the
charactenstlc equatton must necessanly be : m number (countmg
multxphcmes) They are called characterxsnc or latent roots of
the matnx If e desrgnate them by (q. , g.), e can me
11(9) = (q: - q)(q: 9) (q. - a) ==0 (137)
It nsnot easy m fact to nd the charactenstlc roots of a matnx
From the demtxonof a determmant. the coelcnentsof the chame
tenstxc polynomial can be determmed as the sums of pnncnpal
mmors of a g en order (cf Mathematlcal Appendtx A, p 374). and
any of the customary methods for solvmg a polynomlal for tts
roots may then be employed Exen \\ 1th the best short cuts. the
computatxons are tedious Fortunately, the economxst does not
always require the quantitatne \alnes of the roots
Let us return to the problem of ndmg solutions to our reduced
operator system If e put a charactenstlc root. g.. mto the F m
(133), and take for our c's a non vanishing solutnon of (135), ne
shall have a 8011111108
whlch satnses our reduced system Note
that our selectlon olc's ISnot umque, doublmg each wxllalso nge
us a solutxon
A column of c's satlsfymg the homogeneous equatxons xscalled
a charactenshc teclor or latent vectoror sometxmes a modal column
or normal component Some of the charactenstxc roots may be
Frazer Duncan and Collar. ElementaryHolmes chaps w and v
DIFFERENCE EQUATIONS 423
complex; the characteristic vectors will then be complex. Of
course, conjugate complex magnitudes can always be combined so
as to give only real quantities in the nal answer.
If a root g,-is simple, then A(q,)
is not equal to zero, and the
rank of [A giI] is (s -1). Consequently, the latent vector
corresponding to q,-is unique exceptfor scale factor. We may dis
pose of this ambiguity by some normalization convention.
If all roots are distinct, the general solution can be simply
written. Where roots are repeated, certain complications occur.
However, these are not intrinsic and rarely concern the economist.
Later, I show how these difculties can be avoided. But for true
mathematical insight the reader must be referred to the theory of
elementary factors and divisors.
If the s roots are simple, and the 5 latent vectors have been
normalized, we have the following particular solutions of the re
duced equation (131):
CLIFO) gl); 61.2F(t) ga); ''' 61.0F(t1ga);
62.1170: Qi); 02.2170, 92); ' 62.317, Q.);
' ' (138)
m. 9) = %mum. (144)
Because of the property
1;F = 9F, (145)
FD==I+A%+A2tZ!+H-. (151)
For the system
Ey = Ay, (152)
we have
F(t, A, 1; O); = A. (153)
However, for stability investigations these solutions are not very
convenient.
34. If instead of the roots all being simple, some are repeated,
we may write
Mg) = (a: q)"(92 - ) (qr 9). (154)
where (gl, -- -, q,) are the r distinct roots with respective multi
plicities (m1, - ~, m,).
As before, there are special solutions of the form, c,-F(t,qi), but
the c's are no longer unique except for scale factor; in fact, even if
the cs were uniquely dened, there would not be a sufcient num
ber of particular solutions to combine to get a general solution for
s arbitrary initial conditions.
Earlier, in the case of multiple roots we tried solutions of the
F(t. gs) 6F(t. qi) mFU, qi)
, and so we shall here.
form, aq , qz ) ' ': aqua1
Let R(q) be the adjugate of the q matrix,
PGI) = A - QI; (155)
that is, R(g) is the transpose of the matrix made up of the cofactors
of P(g).
By means of algebraic theory of no interest here, it could be
. dkR {
shown 13that from the columns of the matrix [ d? )] can be
selected vectors which when multiplied by the partial derivatives of
F will give a complete solution. However, these difculties can be
" Frazer, Duncan, and Collar, Elementary Matrices, par. 5.10.
426 FOUNDATIONS OF ECONOMIC ANALYSIS
sndestepped by the dev1ceof reducmg the s rst order equations
s dz'erentvanables to stngle sth order equatrons m each vanable "
Thus nf
by ==Ay. (156)
J'o) ="
then
hyUo) = Ab
hyuo) ==Ab
(157)
hy(to) = A "12
Because of the Cayley Hamllton Theorem that every matt-lxsans
es Its own charactenstxc equatlon or
A(A) E 0, (158)
we nd from (156) that
AGM = (zmh'+ uh"! + + my = MAM ==0 (159)
Hence, each vanable say y,(t), satxsesan sth order equation We
also know from (157) the values of [y,(to), hy,(to), , h'y,(to)]
Hence. we have a well determmed system whose solution can be
found by the methods of sectlons 24, 25, and 26 It ls clear then
that our nal soluttons take the form
31'091) "'1-1F(t, Q:)
y,) = 6, MF,Q!)+ cf __" + + 6, liml agmrl
fn,1,
+ C,21170,Q:)+ c, ___
35%:
___q_:) + + cmn,186953.192)
(160)
a r r ""I , '
+ C,"FU' %)+ C,___
ng q ) + + C:m1
a 6958f) v
(J = 1: ,S)
" The method of attack suggested here is not only theoretncally sxmplebut xsalso
computanonally optunal See P A Samuelson Efctent Computauon of the Latent
Vectors of a Matrix Proceedmgs of the Natzonal Academy of Science: XXIX (1943)
397
DIFFERENCE EQUATIONS 427
where the 52csare not all independent. All but 5 are determinable
from the matrix above. The remaining s depend on the : initial
conditions.
Thus, terms of the form Fk/q" may appear if there are multiple
roots. But they need not appear even if there are multiple roots.
That is, whatever the initial condition b, the coefcients c,may
be identically zero for (1 < k < mg). For example, in the matrix
A = E 35,3, (161)
_ man},
* A(h)
(167)
where [T.-;(q)/A(q)] equals [qI -ajl1 and consists of rational functions whose
numerator is of lower degree than denominator. Show that expressions of the
form h"f(t) do not enter into the solution in an essential way.
Its three naturally ordered principal minors, A1, A2, A3, are, re
spectively, ai, 0201_ asao, de(aiz _ dado),and if they are all posi
tive, then the necessary and sufcient conditions for a cubic to be
Routhian are realized. This suggests a general rule for a poly
nomial of the nth degree. List its odd coefcients in a row,
treating all coefcients with subscripts greater than n as zero:
a1a3a5a7.. -. Let each of these elements head a column, whose
subscripts decrease one at each step. Follow the convention that
any negative subscript implies that the coefcient is zero. Thus
we have a square array of numbers, of which only the rst 7:rows
and columns need concern us.
The reader should write out the matrix in full for an 8th and 9th
434 FOUNDATIONS OF ECONOMIC ANALYSIS
degree polynormal, notmg the dxfference between an odd and
even 11
Routh's theorem states that 1!ts necessaryand su'iczentm order
for the real parts of all roots to be negatwe that the prmczpal mmors of
u: momx all be pos-mue, 1e ,
A], = 01 > 0,
Az 0102 " 000: > 0.
(177)
Ail1> 0'
An = anAnI > 0
The condrtlon that all the 0s are posntwe 1scontamed wnthm these
condntzons. but not vxce versa (except 1fall roots are real)
The ngorous proof of thlS theorem cannot be gwen here The
reader must be referred to Rouths Adams Prize Essay A brief
sketch of the reasonmg may, howet er, be 1nd1cated Let X be
any pomt 1nthe complexplane, and f(X ) be a mapping Into another
complex plane Consxder any closed contour C m the X plane
The number of roots of f (X) thhm thts conteur can be determmed
by an Important theorem of Cauchy,23Wthh IS essentlally topo
logical ln nature Traverse the contour C 1n the posxtrve sense
(counter clockwnse) and cons1der the mapped contour of f(X )
The number of roots 111C equals the number of times that f(X)
loops around the ongm , or 1fwe write f (X) as P + Q the number
of tnmesthat the ratxo P/ Q passes through zero from plus to mmus
m excessof the number of tlmes that the ratlo passes through zero
from rmnus to plus
Routh takes as hrs contour C the nght half complex plane. or
more ngorously a lumtlessly large semmcn'cleto the right of the
ongm and the 1magmary was The changes m Slgflof P/ Q on the
sem1c1rcleare easxly evaluated Hence, ll there are to be no roots
m the nght half plane, the number of changes m sxgnof P/Q on
the Imagmary ax1sare uniquely determmed. and the polynomxals
1300 and fg(X) must have all real roots whlch separate each
other, where
i f(zX) = MX) + AMX) (173)
See any textbook on complex vanables
DIFFERENCE EQUATIONS 435
By classical Sturmian theory this involves taking the greatest
common dVSOTOf MX) and f2(X) to get new functions f3(X),
f4(X), - - -. Finally, the conditions on these subsidiary functions
can be shown to be equivalent to the Routh-Hurwitz test deter
minants given in (177).
Numerically, these determinants can be easily evaluated by
various Gauss-Doolittle methods. As an exercise the reader may
verify that the conditions for a fourth degree equation are as
follows:
(11 > 0,
0102 _ 0003 > 0,
010203 ' 00032 _ 01204 > 0,
(179)
a4 > O.
doZn+ 1Znl
+ ' ' + nlZ+ an : as(z _ 1)'(Z + lyv
(185)
__ u-i & _ , azj
takJ k)z (_1) ( ] )Z
where
:: s!
= __' . $ !; I; 0
(t) (s:)!
0. 3<t (186)
ll
l! 0. t<0
DI FFERE.\'CE EQUATIOXS
Collecting terms we nd
(187)
A,.>0,
where these are naturally ordered principal minors of
iii; ; s
do : DI (
OO on Que! NU (189)
~9I9n
.
h
be damped )
Theorem3 If a'a I xsnegatlve demte, thena 15Tmbergentan .
1e . X ,...; = aan stable, and roots of Ia -RI] = are all less than
umty m absolute value (Hmt Xg+1'X+1= X {(du I)X.
= AO: x.) < 0 by hypothesxs Therefore, lug X. = 0)
Theorem4 (Metzler) Il a IS anksxan as dened m chapter
VI , thh all tts o' dxagonal elements posxtwe, then xtxsnecessanly
Tmbergeman
Theorem5 (Metzler) If au 0 and f: a., < 1, then (a I) rs
Jul
Hickman, and hence (by Theorem 4) nsTmbergenlan
Theorem 6 (Mosak) If a 15chkSIan, and ll the o dlagonal
elements of a are all posxtwe. then all of the elements of 01 are
egatzve
DIFFERENCE EQUATIONS 439
Theorem 7 (Oppenheim-Yntema): If all the symbols are posi
tive in